{"collectionById":{"76d23fbd-a202-403b-8313-b3bc36d46679":{"id":"76d23fbd-a202-403b-8313-b3bc36d46679","name":"Research Post","fieldSchemas":[{"id":"1d8c181d-1ccd-446e-87b0-482ccb3ee240","name":"Title","type":"plain_text","role":"primary"},{"id":"571c37b5-f5fc-4742-a5de-b4da6f3415b3","name":"Slug","type":"slug","role":"slug"},{"id":"3fcb25bf-f8b6-47c4-84d6-226369594160","name":"Content (HTML)","type":"rich_text"}],"itemById":{"a54b6460-7b13-40d1-b52f-59d326989a10":{"id":"a54b6460-7b13-40d1-b52f-59d326989a10","index":"#O","collectionId":"76d23fbd-a202-403b-8313-b3bc36d46679","fields":[{"id":"1bee1efa-966f-4518-9cd5-4a6d34550581","value":"OpenClaw is the Signal: Our Thesis on Long Horzion Agents","itemId":"a54b6460-7b13-40d1-b52f-59d326989a10","fieldSchemaId":"1d8c181d-1ccd-446e-87b0-482ccb3ee240"},{"id":"c5fe2891-8ff5-4301-a1b1-a4325a185fc8","value":"{\"root\":{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"For two years, the dominant narrative around AI products centered on a single premise: augmenting humans. OpenClaw breaks that consensus.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Initiated by PSPDFKit founder Peter Steinberger, this open-source project functions as a fully permissioned digital proxy. It reads email, manages calendars, executes code in the terminal, and handles ongoing communications across Slack and Discord. It does not assist. It acts.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"This reflects a structural transition: AI agents are moving from software to labor. The prior generation of SaaS targeted efficiency gains. Long-horizon agents deliver outcomes directly. Pricing logic shifts accordingly — from per-seat to per-outcome, from selling tools to selling labor.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"This memo addresses three questions:\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"What stage does OpenClaw's emergence signal for agent development?\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"As model capability continues to improve, where does the durable moat for long-horizon agents actually sit？\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"As agents move beyond coding workflows into enterprise processes and the physical world, which layer of the stack captures the majority of industrial value?\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"When AI begins to actively execute in the world — rather than simply understand it — a new industrial cycle begins. We are at that inflection point now.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"AI Agents Are Becoming Labor\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h1\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"OpenClaw's emergence marks the first time a new agent architecture has been made viscerally legible to a mass audience: AI no longer just answers questions. It executes tasks over extended periods, operates across systems, and is converging on the profile of a digital employee. This is the long-horizon agent.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Within AI agent discourse, \\\"long-horizon\\\" originates from reinforcement learning's concept of decision horizon: when an agent must traverse extended time and multiple steps before reaching a terminal outcome, classical algorithms break down. Today, as reasoning models and execution frameworks mature, long-cycle task completion is landing in real enterprise environments for the first time.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Technically, the long-horizon agent represents an extended action chain:\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"It decomposes an ambiguous goal into discrete subtasks and maintains state across hours or days.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"It self-corrects continuously during execution rather than running mechanically against a fixed script.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"It handles complex, cross-system, cross-role processes in the real world through deep reasoning and planning.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The more consequential shift is economic. The prior software market was, at its core, a feature-licensing business. When agents begin delivering outcomes directly, the pricing logic, market boundaries, and moat construction of software all undergo structural change. Against this backdrop, three trends are accelerating into formation.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Trend 1: Service-as-Software Drives a Structural TAM Expansion\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h2\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"SaaS maps to roughly $300–400 billion in global enterprise software spend. AI agents unlock the $13 trillion labor expenditure market in the United States alone. That is a \",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"30x TAM expansion\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\".\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Traditional SaaS targets human efficiency gains. Long-horizon agents directly replace full-time employees. This is why outcome-based pricing is being adopted with increasing urgency: customers no longer pay for features — they pay per resolved ticket, per completed workflow, or per unit of labor cost eliminated. The software business model is migrating from selling tools to selling labor. Players such as Sierra and Decagon are already pricing per outcome — successful resolutions per conversation — and tying take rate directly to the labor cost savings realized by the customer.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"We are at an inflection point in agent unit economics. Through 2024 and into 2025, most agent companies have looked more like labor outsourcing firms than software businesses. Expensive underlying model costs — COGS running approximately 70% — combined with intensive human-in-the-loop intervention have compressed gross margins to 40–50%. Long-horizon tasks compound the problem, driving token consumption to scale nonlinearly.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Two structural changes are driving the inflection:\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"First: logarithmic collapse in inference cost.\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\" Inference cost drops an order of magnitude every 18 months. The unit economics of individual tasks are beginning to invert.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Second: the proliferation of reasoning orchestrators and tiered model dispatch.\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\" A growing share of agents route complex planning to high-capability models while delegating execution steps to cheaper models, materially restructuring the cost stack.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Agent unit economics may rapidly converge back toward software-product margins, executing a full-dimensional displacement of traditional services businesses.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Trend 2: From System of Record to System of Action — The Moat Shifts\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h2\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The core of the prior generation of enterprise software was recording the world. Systems of record such as Salesforce depend on users continuously inputting data — precisely the activity enterprises are least willing to perform. Long-horizon agents represent a \",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"system of action\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\": software no longer merely records; it executes directly.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"This shift introduces a new moat logic: \",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"workflow data gravity\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\".\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"If Stripe's defensibility derives from payment data, agent defensibility begins to derive from execution traces. Every task run accumulates corner cases, human correction records, and API call paths. This data does not appear in public training sets, yet it materially improves accuracy within specific enterprise environments. Agents fine-tuned on proprietary data consistently outperform general-purpose models in vertical deployments. Switching costs become prohibitive — a generic model cannot replicate an agent that has been calibrated through real operational history.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Computer-use agents such as Simular operate directly at the OS layer, converting previously unstructured mouse and keyboard behavior into learnable execution traces. This progressively builds a data barrier that is difficult to replicate.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"This raises a pointed question: as model capability continues to compound, what is actually scarce — intelligence or experience?\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Trend 3: High Agency and Voice — When Agents Begin to Work Proactively\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h2\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The interaction paradigm of 2026 shifts from passive response to proactive intervention. The next-generation agent does not wait for instructions. It continuously monitors its environment, surfaces recommendations, and executes autonomously once authorized. In its ideal state, it operates like an S-tier employee: identifies the problem, runs the diagnosis, develops the plan, executes the work, and requests approval only at the end.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Consider the current sales workflow: reps manually update the CRM. The future agent automatically analyzes two years of email history, surfaces dormant accounts, and drafts follow-up outreach. The user clicks approve. Sales cycle compresses materially.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"In this context, voice is transitioning from an interaction interface into the face of the digital employee. In high-trust verticals — healthcare, insurance, financial services — voice is not merely UI. It is a critical carrier of compliance and emotional management. Voice agents demonstrably outperform humans on compliance adherence because they do not commit the inadvertent violations humans do. End-to-end voice reasoning models now enable agents to interpret emotional state in real time, navigate complex conversations, and close end-to-end workflows. Some companies deliberately introduce latency and ambient noise to close the psychological distance between AI voice and human voice.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Voice agent plus proactive agent is likely the defining form factor of the AI employee.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"2026 Agent Investment Thesis\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h1\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The central question for agents in 2026 is where value accretes in the stack. We focus on companies leveraging long-horizon capability to solve high-value commercial problems.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Over the past two years, capital has been most concentrated in coding agents. It is a closed, high-determinism environment — the easiest starting point for agents to prove out. But as encapsulation capability improves, the real opportunity is migrating from the code world into enterprise processes and live business operations.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"From an investment perspective, we track companies across four theses.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Thesis 01 — Reasoning Orchestrators: Infrastructure That Lets Agents Work Longer\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h2\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Agents require state management. In many cases, agent task failure is not a model intelligence problem — it is a failure to maintain state across a long-running execution. Long-horizon agents are long-lifecycle software. They must execute across hours or days and run continuously across asynchronous systems. Durable execution and state management are emerging as a new infrastructure layer.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Companies such as \",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Temporal\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\" and \",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Inngest\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\" ensure that if an agent fails at step three and the server goes down, execution resumes from step three on restart rather than from the beginning. Their durable execution guarantees state persistence. \",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Parallel Web Systems\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\" represents a distinct infrastructure direction: rebuilding the internet itself for agents. The traditional web was designed for humans. Agents require a predictable, noise-free execution environment. As agents become the primary users of the internet, they need a clean environment without popups or CSS clutter. The agent-first web may become the new traffic entry point.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Thesis 02 — Process Intelligence: Model Capability Converges; Execution Experience Does Not\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h2\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h2\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Foundation models have absorbed the majority of public data. The remaining value is embedded in enterprise-internal processes, logs, and employee behavior traces. The significance of execution traces lies in capturing process data. Whoever can record which three documents a claims adjuster reviewed and where they paused before rejecting a policy holds a vertical model moat. We focus on companies that learn from human expert experience during execution — capturing the path to problem resolution, not just the resolution itself.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Companies such as \",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Simular\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\" have accumulated substantial visual-action data from operating legacy software — EMR systems, legacy banking platforms — in financial and healthcare settings. This is terrain general-purpose models cannot easily access. The JPMorgan commercial lending case is instructive: the workflow is not conversational. It involves reading screens and filling forms. \",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Mimica\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\" observes employee screens and records the trace of seeing error A, clicking button B, copying code C — converting tacit employee knowledge into explicit code. These traces are translated directly into agent execution logic. Compared to traditional RPA, what Mimica captures is the human decision path.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Thesis 03 — Selling Labor: Real Agent Companies Sell Outcomes\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h2\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h2\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"A simple test for whether a company has genuine agent potential: is the customer paying for software, or paying to not hire a person? We prioritize companies willing to price per FTE or per outcome. Gross margins may look compressed early due to token consumption, but the displacement potential is strong. This space contains many vertical specialists achieving meaningfully higher accuracy in specific high-value verticals — insurance, legal, procurement.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Serval\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\" is the canonical example. Customers pay because it directly replaces IT support headcount. When an engineer requests access permissions in Slack, the agent completes the approval, execution, and audit loop in seconds. Customers such as Verkada anchor the value proposition explicitly to FTE savings.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Distyl AI\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\" operates on a high-touch delivery plus platform accumulation model. Early-stage forward deployed engineers embed deeply in customer operations. Once the pattern stabilizes, it produces a high-margin service-as-software structure. It resembles Palantir for the AI era, but with faster delivery cycles — executing bill forecasting for T-Mobile and medical prior authorization for Elevance Health.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"In vertical markets, companies such as \",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"WithCoverage\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\", \",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Corgi\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\", and \",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Omnea\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\" apply long-horizon capability to insurance auditing and procurement workflows. These tasks are low-frequency, high-value, and high-complexity — the scenarios where agent displacement potential is strongest.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Thesis 04 — Voice Agents: The Face of Labor\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h2\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h2\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"In the selling labor context, voice is the agent's interface to the real world. If the reasoning orchestrator is the brain and the system of action is the hands, voice is the digital face that engages human customers directly, manages complex emotional states, and closes end-to-end workflows.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Leading voice agents in 2026 no longer run traditional STT-LLM-TTS pipelines. They use end-to-end native audio reasoning, compressing latency below 300ms with full interrupt support. Many long-horizon tasks are also emotionally driven — customer anxiety and frustration on a claims call, for instance. Agents capable of handling emotional context achieve materially higher task completion rates than conventional chatbots.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"A Representative Long-Horizon Agent Stack\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h1\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The four theses above map cleanly onto a full stack, from base infrastructure to final delivery. \",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"type\":\"image\",\"version\":1,\"hash\":\"9c88f83e6ab32254e57c4b3120b9ae0f0611396f\",\"src\":\"https://s3-alpha-sig.figma.com/img/9c88/f83e/6ab32254e57c4b3120b9ae0f0611396f?Expires=1774224000\u0026Key-Pair-Id=APKAQ4GOSFWCW27IBOMQ\u0026Signature=qLSpvb6s7TjTZrasNUFeipsAQf43gbq4Byx4stCovE07ghG9qmusJdTLfPZtz9oNCkl7-v~VTtiOG36uIY7bjFUqXskzHXn5SXNiYEnVg8-hQbupF-gZ0dPELkraafnrFrq0qYOXf~si6fkNzwJ-tp92fbnD-nE0vr0glUJTA0KPB22WGCMKD0JuzKNw559d5ksmv~SmoAKpH007mJgl1b8fm9s9AAI4XOQxcIuhdfpCf3~vaV2TueTE70RwgwcNy~PgrQfMizItAo8K9Q1U~aoJGzhBw5zHs5CLxa7hMnBOFSqdp9U1EdIh7wcfO6XX4KP3EuPEUR9vdcFLAuCFmw__\",\"altText\":\"\",\"originalImageWidth\":2500,\"originalImageHeight\":2337,\"isFillWidth\":false}],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Take insurance as a worked example. A single claims workflow decomposes across the full stack:\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"LayerCompanyRole\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Interface—ElevenLabs / Retell AI.\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"A customer calls in from an accident scene, acutely distressed. The agent responds within milliseconds — calm, empathetic — de-escalates the customer and extracts the key facts.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Brain—Distyl / Custom Model.\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The backend model takes the extracted facts, retrieves the customer's policy terms via RAG, and reasons against local traffic regulations.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Eyes and Hands—Simular.\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The model determines it needs to access the claims system — a legacy platform built 20 years ago with no API. The agent spins up a virtual environment, operates the mouse to log into the backend, completes the approval, and generates the PDF.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Safety Net—Temporal.\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The system crashes mid-PDF generation. Temporal catches the failure, automatically retries, and ensures the workflow continues without human intervention.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Evolution—Mimica.\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The entire workflow is recorded as an execution trace. If the agent handled it well, that trace is automatically added to the fine-tuning dataset tonight. Tomorrow, the agent is more capable.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Long-Horizon Agent Landscape\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h1\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Mapped against task environment closure and value delivery layer, the agent market segments as follows. Each domain has the potential to produce multiple unicorns.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"type\":\"image\",\"version\":1,\"hash\":\"e4eef1701a7b47312d8a57732b11568ee0959ad4\",\"src\":\"https://s3-alpha-sig.figma.com/img/e4ee/f170/1a7b47312d8a57732b11568ee0959ad4?Expires=1774224000\u0026Key-Pair-Id=APKAQ4GOSFWCW27IBOMQ\u0026Signature=nyHp-NjAK0B8dBHoDvkF3P9uoJRik~mcY9oT2d93k3gDeE7a-X5vJENwXwlBnyDWm63nq9QLDhZmHgKjQhPXYeP89QiZWkCwAkYxRyS3vSZwYHuz0XjHlpqBhbp0o0teFgf~JxW7pCqFSOVQ5pkjQQmXy25ecbostJQIfNm71szvJW53wtvfN5lJkC8EujGkVAoPJCbJS3TMsHWXGza0rV3FJ5agWZwOy1efnvrcGU1LJZJf1WsrEUKUuuvKFMw6kkt8p0cjReAUW--6T1QGjQtkpnXIktpAMJDhPzoA2mNiilEoACwpGsVcEiM5cNDbARI4ZySWU67YWT8aTHiAHg__\",\"altText\":\"\",\"originalImageWidth\":4688,\"originalImageHeight\":3750,\"isFillWidth\":false}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"X-axis:\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\" Task Environment — from purely code and digital environments → enterprise workflows → physical world and complex environments\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Y-axis:\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\" Value Delivery Layer — from infrastructure → platform tools → end-to-end services\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":2}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The Horizontal Axis: Environment Complexity\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h2\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Left side — code and structured data\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\" is a pure logic world. The environment is closed, feedback is deterministic: code either runs or throws an error. This is where agents broke out first and where capital has been most concentrated over the past two years.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Center — GUI, SaaS, documents\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\" is the gray zone of enterprise process. It combines APIs, documents, and mouse-driven interactions. Rules are not fully explicit, but commercial value is high. This is the primary battleground where selling labor becomes a viable business model.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Right side — open internet, physical world\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\" represents real-world complexity. Popups, anti-scraping mechanisms, voice emotion, and unstructured interaction coexist. This is the domain of voice agents and open-web agents: the highest technical difficulty, but potentially the largest long-term distribution surface.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The Vertical Axis: Value Delivery Layer\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h2\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Bottom layer — infrastructure and tooling\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\" provides durable execution, connectivity, and stability. This is the picks-and-shovels layer.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Middle layer — OS and models\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\" is the agent's brain and operating system, responsible for reasoning and task orchestration.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Top layer — application and platform\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\" is the selling labor layer: directly replacing human labor, priced per outcome, and commanding the highest enterprise budget allocation.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"This map reflects a value migration already underway in the agent ecosystem. Over the past two years, capital concentrated on the left side of the map — software engineering — a closed, highly structured code environment that has become a red ocean. In 2026, the center of gravity shifts to the middle axis: the mixed GUI and SaaS environment, moderate noise, highest commercial value.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Software Engineering\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h2\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Coding agents were the first long-horizon agent deployment to prove out at scale. The environment is closed, feedback is deterministic. As a result, agents moved from copilot to autopilot here first — taking over the full chain of write, run, debug, test, and deploy.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"From an investment perspective, however, this vertical carries a structural tension: \",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"it is the easiest domain in which to demonstrate value, and the easiest domain in which encapsulation erodes differentiation.\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\" Breakout growth happens here; margin and moat do not necessarily stay here. We therefore segment software engineering into three categories, each with a distinct winner logic.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"type\":\"image\",\"version\":1,\"hash\":\"33f0c48632fcba99b9b8a81672e74d2d2da08d3e\",\"src\":\"https://s3-alpha-sig.figma.com/img/33f0/c486/32fcba99b9b8a81672e74d2d2da08d3e?Expires=1774224000\u0026Key-Pair-Id=APKAQ4GOSFWCW27IBOMQ\u0026Signature=UqCRql3JokSFhFjdZsO6BBpwlKl2uA-eDf3S~nwxhbmouFhxThL4ThLxHIlMYgcPRl1epZ18PytvA1vE6zg-g0jYgeLyPvOBAOlIUGOFfOo2e230-7CaPEOxQiQITy9eIto3RyBMrL60YZU69jnmq3S7LV~8tKxHFIZn0NkooZZKYVwXpAGke1UlNMDATii2aQd1Nnj4MY04PBJrjtDKa5dAfZPwRj~nt2Ir6ut6GUpgUnAVrrmUqGbMF28l1riEpavlswmwzwRbt35ZpbqXhroTWQ2xN6paDp2po7xfyeX3N71cT3LLfmUdB77OKQpKLBQe7rM9~7GsLmn2g9mxIg__\",\"altText\":\"\",\"originalImageWidth\":1014,\"originalImageHeight\":1004,\"isFillWidth\":false}],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"1) Vibe Coding: Turning Software Development into Demand Expression. The ceiling is determined by reliability and distribution.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h3\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Demand-side growth is outpacing the supply-side increment of engineers. A growing base of founders, operators, and salespeople are building their own tools — this is the core driver behind the rapid emergence of vibe coding products. But once these products enter core enterprise systems, reliability, permission management, audit capability, and cost control rapidly become the binding constraints. Future winners will either control a strong distribution entry point — an IDE or platform-level surface — or productize controllability itself: version management, testing mechanisms, permission architecture, and rollback capability.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Representative company: Emergent\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The Emergent team combines the co-founder of Dunzo with core members from Amazon SageMaker — clear technical and commercialization DNA. The company is backed by Lightspeed, Google, and SoftBank Vision Fund 2. The product reached $15M ARR within 90 days of launch and $50M ARR at seven months, demonstrating strong early growth momentum. The user base is primarily developers, SMBs, and founders — widely adopted by non-technical founders building MVPs. Current use cases remain concentrated in SMB internal tooling and prototyping; long-term moat and production-grade reliability are yet to be validated.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"2) End-to-End Agent: From Writing Code to Delivering Outcomes\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h3\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"This category targets the full workflow a real engineer executes: task decomposition, environment configuration, debugging, testing, PR submission, and deployment.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Two distinct strategic routes are emerging.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"IDE inside-out — Cursor, Replit — starts from the developer entry point, embeds agents into daily workflow, and trades human oversight for reliability.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Agent outside-in — Devin, OpenHands — positions as a digital engineer, emphasizing greater autonomy and end-to-end delivery.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Representative companies: Cursor, Replit, OpenHands\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Cursor’s core value is an agent-native development workflow entry point.\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\" Through Composer, it upgrades the code editor into a semi-autonomous development environment capable of cross-file understanding, terminal operations, and multi-step planning. Relative to fully automated approaches, Cursor bets on long-chain human-in-the-loop collaboration — a positioning that has produced materially faster adoption in real production environments.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Replit has shifted its strategy from online IDE to an AI app builder spanning idea to deployment.\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\" Through Replit Agent, it integrates environment configuration, frontend and backend development, database management, and cloud deployment into a single continuous workflow, lowering the barrier for non-professional developers to ship applications. Its differentiation lies in distribution advantages rooted in a strong community and education heritage, driving a prosumer-to-SMB bottom-up penetration path. Relative to Cursor’s developer-first positioning, Replit is closer to an application generation platform — an attempt to become the low-barrier software production infrastructure of the AI era.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"OpenHands is pushing the autonomous software engineering agent from demo toward enterprise-grade infrastructure.\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\" Its notable capability is scaling from a single agent to thousands of agents operating concurrently, with emphasis on genuine task execution in sandboxed environments — writing code, running CLI commands, browsing the web — rather than generating suggestions. A partnership with AMD reinforces edge inference and cost structure advantages. The team combines Google engineering backgrounds with top-tier academic AI researchers, giving it strong technical credibility within the developer-native ecosystem.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"3) Remediation: When Code Generation Cost Approaches Zero, Maintenance Becomes the New Labor Budget\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h3\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"As AI-generated code grows exponentially, the cost of maintaining code and ensuring system stability is rising sharply. Remediation is where selling labor lands in the developer tools domain: AI SRE, automated repair, code review, migration, and technical debt clearance. Within software engineering, remediation carries stronger payment resilience — it maps directly to enterprise stability budgets, not developer tooling budgets.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Representative companies\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Resolve and Traversal represent the AI SRE direction.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Resolve introduces AI agents into production engineering, upgrading the traditional SRE function from reactive alert handling to a multi-agent autonomous operations system. Its core capability is cross-context reasoning across code, infrastructure, and telemetry data — automatically triaging alerts, identifying root causes, and generating repair code with production environment context, compressing MTTR. The founding team comes from the core observability layer at Splunk, layered with OpenTelemetry ecosystem background, positioning Resolve as an extension of the observability-to-autonomous-ops trajectory.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Traversal’s differentiation is upgrading AI SRE from simple automation to complex system analysis grounded in causal reasoning. It deploys parallel agent clusters to run statistical tests and causal modeling, cutting through log and metric noise to isolate deep system issues, and can execute remediation plans autonomously. Relative to traditional observability tooling, it emphasizes zero vendor lock-in and cross-system analysis, positioning as an intelligent orchestration layer for multi-cloud environments. The team leans academic AI and reinforcement learning, backed by Sequoia and Kleiner Perkins — an infra-native research-oriented company betting on complex system debugging as a high-barrier technical direction.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"CodeRabbit and Sweep are closer to outsourcing repetitive maintenance to AI, most effective on low-to-medium complexity tasks.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Sweep AI focuses on ticket-level autonomous coding, positioned as a junior developer proxy. Through a fire-and-forget model, it automatically reads the codebase, generates PRs, and self-iterates based on CI/CD feedback until tests pass. Its differentiation is a deliberate focus on high-frequency, low-creativity GitHub Issue workflows — advancing AI from coding assistance to executable task layer and reducing the time senior engineers spend on routine work. The product performs reliably on well-defined small tasks but remains constrained on complex architectural reasoning, operating closer to an automated execution layer within the development workflow.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Enterprise Action Systems\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h2\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Enterprise Action Systems are defined by agents that\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\", within fragmented, nonlinear, exception-heavy enterprise environments, do not stop at recommendations and summaries. \",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"They take over the full execution loop\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\" — intent recognition, context retrieval, permission and compliance validation, system execution, audit and rollback generation — and directly replace human labor.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"We believe this market becomes the primary battleground in 2026, for three reasons.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"First, the value anchor is unambiguous: savings are denominated in FTE, BPO, and ticket costs.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Second, the environment is complex but bounded — less noisy than the open internet, but closer to real business operations than a pure code world.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Third, high deployment barriers translate into thicker moats: permission management, audit trails, integrations, staged rollouts, rollback capability, and exception handling are all required. Point-in-time model capability cannot substitute for this stack.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"link\",\"version\":1,\"rel\":\"\",\"target\":\"_blank\",\"title\":null,\"url\":\"https://substackcdn.com/image/fetch/$s_!twj-!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fffaf3e62-8923-4d21-90fd-7a4d90d943e9_2500x1925.png\"}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"type\":\"image\",\"version\":1,\"hash\":\"1d1546839dd41941e1971fa2f67a810c61f37009\",\"src\":\"https://s3-alpha-sig.figma.com/img/1d15/4683/9dd41941e1971fa2f67a810c61f37009?Expires=1774224000\u0026Key-Pair-Id=APKAQ4GOSFWCW27IBOMQ\u0026Signature=HoG95-GPoQ4r0iPwbTeTDBBe-UGzElsDHsClSFmfHVbwKCSJ2Fr4o~orQof3X7CXPxt9NzAhmJQgfClFwN~7cCiXC7TYZeMctmO46Fsa87Sg6hN9wUxQ0TfKaGliIyYFQ6vp6zpm1WnD5pZdqEgTd608jj85tVWyLnyPCv2nas-vnvCxA5fE3kHkK7RAbCoMt3LdLRSt93TZr9fPAfFwSxO01zjmtg5a73FtLZltf65xv88DPqwk0tldBbl3WwZwjw1daIuN2~F3CQee-xzzsEf~uB7icRj4n1MQDDJRO-p1TqdJAOfDkNt5h9696OU3aIi2-rGSX2wc0xwvo-icnA__\",\"altText\":\"\",\"originalImageWidth\":2500,\"originalImageHeight\":2487,\"isFillWidth\":false}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"1) Horizontal Ops: Enterprise Execution Layer — Enter Through One Department, Expand to Everything Ops\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h3\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"These companies typically enter through IT, HR, or finance ticketing and workflow, deploy a unified agent execution layer across multiple departments, and target becoming the internal action hub of the enterprise.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Representative companies:\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Serval upgrades ITSM from a system of record to an AI-native system of action. Users describe needs in natural language via Slack; the agent generates and executes auditable workflow code, closing the loop on permission management and approvals. Revenue grew explosively within 18 months, and multiple customers have decommissioned their legacy ITSM systems — validating that Serval has timed the generational transition in enterprise ops automation. Long-term moat depends on whether it can become the system of record for agents.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Ema productizes AI agents as a Universal AI Employee, deploying persona-based digital workers — customer service, documentation, data science — so enterprises can staff AI the way they hire people. Its Generative Workflow Engine and multi-model architecture emphasize long-workflow task decomposition and compliance reliability, achieving 70%+ automation rates and measurable ROI in financial services, customer service, and HR. Ema bets on outcome-oriented pricing and role-based delivery, targeting the AI workforce layer inside the enterprise.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Relevance AI positions as a low-code agent OS, focused on building and orchestrating multi-agent teams, helping enterprises decompose complex workflows into continuously running long-horizon task systems. Its differentiation is orchestration and visual operations — suited for non-engineering-heavy teams looking to stand up AI automation quickly.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"2) Custom Delivery: High-Touch Delivery Plus Platform Accumulation — Solving Data Silos and Complex Workflows\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h3\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Deploying AI agents into core enterprise processes requires engineering capability to bind data, systems, people, and compliance into a closed loop. These companies advance Palantir-style via forward deployed engineering: use delivery to solve the cold start problem, then accumulate repeatable patterns into platform components.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Representative companies:\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Distyl AI was founded by former Palantir members Arjun Prakash and Derek Ho, positioned as a customized agent build-and-delivery platform for high-complexity enterprise workflows. It embeds large models into Fortune 500 core business processes through a strategic partnership with OpenAI. The model is closer to a consulting-SaaS hybrid: early-stage high-touch engineering teams resolve data silo and compliance problems, which are then progressively distilled into reusable platform components. Current advantage lies in rapid deployment in complex scenarios that traditional SaaS cannot reach — supply chain forecasting, billing anomaly detection. Long-term scalability depends on whether the delivery model can transition from project-based to productized.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Mimica was founded by former Palantir engineer Tuhin Shroff and AI researcher Raphael Holca-Lamarre. It focuses on task mining and automated workflow generation — observing employee operations in the background to discover complex processes automatically, then generating RPA scripts or standard process documentation. Customers are concentrated in high-compliance traditional industries: Liberty Mutual, Goodyear, Merck, DHL. Its value is in solving the hardest phase of enterprise automation — process discovery — and delivering fast, demonstrable cost savings. The product currently sits closer to the efficiency optimization layer; whether it can become a long-term system entry point requires further observation.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"3) Vertical Specialist: Replacing FTEs in High-Compliance, High-Value, Domain-Intensive Scenarios\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h3\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"When processes and risk profiles are deeply industry-specific — insurance, healthcare, procurement, finance — horizontal tools struggle to absorb the necessary detail. Vertical companies more readily achieve end-to-end role replacement in a single function and command higher outcome-based pricing.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Representative companies:\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Finance and Procurement: Sema4.ai (finance document workflows), Omnea (procurement and vendor governance)\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Sema4.ai is driven by former Cloudera and Docker CEO Rob Bearden alongside the Robocorp founding team, positioned as an enterprise-grade agent execution platform with a focus on long-workflow finance and compliance tasks. Its technical approach combines Rasa’s conversational understanding with Robocorp’s Python automation, enabling agents to directly execute runbooks and actions rather than remaining at the recommendation layer. Koch Industries is an existing customer, deploying it for invoice automation.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Omnea was founded by former Tessian executives Ben Freeman and Ben Allen, focused on AI-driven ingestion and orchestration of the source-to-pay procurement workflow, targeting a unified entry point for enterprise supplier data and risk management. Customers include Spotify and MongoDB. Its differentiation is integrating procurement intake, contract renewal, and risk governance into a single system. Current value is more concentrated in workflow efficiency and data centralization; whether it becomes a long-term replacement for the procurement system of record requires more time to validate.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Insurance: WithCoverage (broker replacement), Corgi (AI-native carrier), Further AI (underwriting and claims automation)\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"WithCoverage was founded by Opendoor co-founder JD Ross and former Bain and Compound team member Max Brenner, positioned as an AI-native risk management platform replacing traditional insurance brokers. Through a flat-fee model and an AI risk audit engine, it provides insurance portfolio optimization and claims management to growth-stage companies — converting commission-driven brokerage into a more transparent software-based service. Customers are concentrated in high-growth technology companies. Core advantage lies in cost optimization and process digitization, but the company remains structurally a broker-layer business; moat derives primarily from customer relationships and data accumulation.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Corgi is an AI-native insurance carrier — it designs products directly and bears underwriting risk. Founders Nico Laqua and Emily Yuan bring startup and operations backgrounds. The team is approximately 70 people, backed by Y Combinator and Kindred Ventures, with total funding of approximately $108M. The product emphasizes end-to-end automation from underwriting to claims, with modular coverage designed around startup financing stages — D\u0026O, cyber, AI liability. Relative to broker-model companies, the critical variable is whether its risk management and pricing models can sustainably outperform incumbent insurers.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Computer Use \u0026 Prosumer\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h2\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"OpenClaw gave the mass market its first visceral sense of high-autonomy agents. Computer use — reading screens and operating a mouse — is the most critical underlying capability, and one of the most significant technical bottlenecks today. From an investment perspective, computer use is best understood as the mechanism that lets agents penetrate the 80% of enterprise systems with no API — legacy infrastructure and fragmented processes — while giving individual users a pair of outsourceable digital hands.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Computer use means executing long-horizon tasks in high-noise environments: popups, anti-scraping mechanisms, page redesigns, permission gates, CAPTCHAs, network instability, and state transitions across collaborative software. This creates three structural constraints.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Reliability thresholds are extremely high:\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\" in any long-chain task, a single step failure collapses the entire workflow.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Cost structures are more sensitive:\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\" visual understanding combined with multi-round retries drives token and inference cost to scale nonlinearly.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Security and compliance are hard gates:\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\" the closer an agent gets to hands-on execution, the more it requires an auditable, rollback-capable, permissioned control plane.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"type\":\"image\",\"version\":1,\"hash\":\"375349c97f62a7de60eb5ef13ff3bbbed0826e83\",\"src\":\"https://s3-alpha-sig.figma.com/img/3753/49c9/7f62a7de60eb5ef13ff3bbbed0826e83?Expires=1774224000\u0026Key-Pair-Id=APKAQ4GOSFWCW27IBOMQ\u0026Signature=IKTaZnPKS2fciXiI6VV024tzQnFcZkEPApqJcCqxqbUeat3Wgvkgjb8JTEUreBy3AgY6UN9q3ws9pwg~-3cPp4y8DlMePluRSK24KXlKTFhKMmYxwMwMHq4QFzTKUOU~z~~FslzQxO2kLmb2DNwmXUyMaOws3AXuuIKqbDpQyuhuCfdlEJIi~hU4nJ4z-lxZjkPKG66zgVZbpRsHi9lo3mvS5GWMr-wgE0fHd9CYHnmMHg6s02HCDFkI15VETy1KHoQr6juHbcg1gVqZ46lfZ75C1W1rb-eWJcStm0fN6QmYAD6EUK0aWG6dp0kPig8XBD0HXLNcmfDhEXnSASx48Q__\",\"altText\":\"\",\"originalImageWidth\":2500,\"originalImageHeight\":2487,\"isFillWidth\":false}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"center\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"1) OS Level: System-Layer Takeover — High Value, Heavy Delivery\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h3\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h3\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"OS-level agents focus on operating local applications and enterprise legacy systems — Excel, SAP, EHR, Citrix, and assorted legacy backends — targeting the last mile of enterprise process. Integration, permission, and delivery complexity are correspondingly higher.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Representative companies:\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Simular was founded by former Google DeepMind researchers Ang Li and Jiachen Yang, positioned as a local computer-use agent operating desktop software — Excel, SAP, EHR — directly on Mac and Windows. The company emphasizes on-device and controlled execution, supports long-chain tasks through its Agent S2 architecture and memory mechanisms, and has joined Microsoft’s Windows 365 for Agents program. Its trajectory leans toward enterprise IT automation rather than consumer assistant.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Manus positions as a general-purpose autonomous agent, emphasizing end-to-end execution capability from research to output — representative of the execution layer route. The company had Benchmark backing prior to its acquisition by Meta, with a strategy of minimizing human intervention on complex tasks. As foundation model providers increasingly internalize agentic capability, general-purpose execution layer products face growing platform competition pressure.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"2) Browser Level / Web Automation: Faster Distribution, but Vulnerable to Platform Rules\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h3\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Browser-level agents are lighter, more accessible to prosumers, and more amenable to subscription pricing. The primary risk comes from the open internet itself: anti-scraping, CAPTCHAs, site redesigns, authentication state, payment flows, and fraud controls.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Representative companies:\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Yutori was founded by a team of former Meta FAIR researchers, building autonomous web agents for cross-site monitoring and automated operations — price tracking, form submission, and similar workflows. The company raised a $15M seed from Radical Ventures and Felicis. Its product Scouts emphasizes long-cycle web tasks and reliable execution, currently in beta. Strength lies in the research team and web-native agent architecture; realized value remains dependent on task stability and well-defined prompt design.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Twin was founded by former Dreem founders Hugo Mercier and Joao Justi, focused on headless browser automation. It uses visual and action models to operate web UIs autonomously, attempting to upgrade traditional RPA into AI-native web automation.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Infrastructure\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h2\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Three practical questions determine whether an agent can actually deploy at enterprise scale: \",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"can it run long enough, can it run reliably, and can it run under control\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\". Long-horizon agents must span hours or days, traverse asynchronous systems, handle failure and retry, maintain audit trails, and support rollback and human takeover at critical junctures. All of this must be built from infrastructure up.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"In the landscape, infrastructure is the foundation that determines whether upper-layer applications can move from demo to scaled delivery.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"type\":\"image\",\"version\":1,\"hash\":\"d11b65b1b8f0cc5441139d3ce0e2c3a1eeead7aa\",\"src\":\"https://s3-alpha-sig.figma.com/img/d11b/65b1/b8f0cc5441139d3ce0e2c3a1eeead7aa?Expires=1774224000\u0026Key-Pair-Id=APKAQ4GOSFWCW27IBOMQ\u0026Signature=moCCiVjsOEmYEahlug2FwFDsuu-czP-fLSTiJ~quwevT9ldFTEkAtKCD3cJ4GuUdy18Lkwp84KxQOM8Gfr-oyFXvP38o6oOoj8oyiKFGd8mUD4giajY0Z~8NyITlM4M6XCOAd7JdSoNsHpkrURSEqmOpsqhRz6FrpZ32q91OxDTp3nFqptHdFVH-vL6F8G8WTb2egZQcKraUDtq0NFfUKDc28aS0FXfnisMgwtiYyBZol7lpY5DQJ0RWmGtVUkATlpp5iaBnnsYBErkvBac-jtatALpjqLepDtL88fKmdBG7HE~-hesf6KRb8weFbWXPXCUCiz9YobAG82sHnt0bvg__\",\"altText\":\"\",\"originalImageWidth\":800,\"originalImageHeight\":796,\"isFillWidth\":false}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"1) Agent-First Web and Environment\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h3\"},{\"children\":[],\"direction\":null,\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h3\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The first step for a long-horizon agent is information retrieval — but the real internet is deeply hostile to agents: noise, anti-scraping, popups, structural chaos, and closed content. This layer converts an uncontrollable web into an executable environment.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Representative company:\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Parallel Web Systems represents the web-for-AI-agents route. It rebuilds the internet as an agent-usable execution surface: more stable scraping and search APIs, more controllable interactions, stronger provenance and confidence mechanisms. Its products are agent-first web browsing and search infrastructure — providing AI agents with search, scraping, and interaction APIs distinct from the human web, solving for agent usability and stability in real web environments.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"2) Workflow Orchestration\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h3\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Long-horizon tasks require cross-system state management, retry strategies, idempotency, and observability. Companies such as Temporal, Inngest, and LangChain provide durable execution infrastructure, solving the core problems of state persistence, failure handling, and automatic retry across extended agent tasks. Traditional automation incumbents such as UiPath and Zapier are leveraging their large API connector ecosystems to transition from rule-based linear workflows toward dynamic AI agent orchestration.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Representative companies:\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Temporal was founded by Maxim Fateev and Samar Abbas. It provides a durable execution engine for long-running business processes, addressing the foundational problems of state management and failure recovery in distributed systems. Relative to traditional workflow tooling, it operates closer to the infrastructure layer: developers write business logic as recoverable state machines via SDK, and the engine guarantees task continuation after restarts, timeouts, or service failures.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Inngest positions as a lighter-weight workflow orchestration platform, emphasizing a durable-by-default, event-driven execution model. Relative to Temporal, it reduces infrastructure complexity — prioritizing serverless and runtime-agnostic architecture — enabling developers to rapidly add retry, recovery, and observability to AI workflows, background tasks, or agent orchestration. Its advantage is low onboarding friction and alignment with modern development stacks. Long-term differentiation depends on whether it can build a developer ecosystem in the AI orchestration space.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"3) Model as Agent\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h3\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Internalizing deep reasoning, code generation, and computer use capability into model weights, building agents capable of directly operating browsers, terminals, or desktops.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Representative companies:\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Imbue was founded by Kanjun Qiu and Josh Albrecht, positioned closer to a research-oriented AI lab. Its core bet is reasoning capability as a long-term moat. The company has developed the Imbue 70B reasoning foundation model, explores multi-agent collaborative coding environments through the Sculptor interface, and open-sources training infrastructure tooling such as CARBS — emphasizing a full-stack research path from model to execution environment. Backed by Nvidia, Astera, and Eric Schmidt, with a large H100 cluster, the strategy prioritizes research depth over near-term commercialization, competing with OpenAI and Anthropic at the foundational research layer.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Reflection AI was founded by former DeepMind researchers Misha Laskin and Ioannis Antonoglou, focused on long-chain reasoning models with reflection mechanisms, targeting improved autonomous coding and high-autonomy software execution capability. The company has raised substantial funding led by Nvidia, totaling approximately $2.13B at an $8B valuation. Similar to large model companies, its strategy explores agent application capability upward while internalizing tool use and workflow orchestration into model weights downward. The company remains in a research-forward stage; the commercialization path is not yet fully defined.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Voice Agents\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h2\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Progress in voice agents is measured across three dimensions: real-time responsiveness, emotional handling, and controllability.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Leading voice agents in 2026 are moving away from the assembled STT-LLM-TTS pipeline toward more end-to-end, lower-latency real-time architectures. End-to-end response below 300ms closes the gap with human conversation rhythm, reducing interruptions, topic drift, and call abandonment. Barge-in capability — the ability to stop and redirect the moment a customer says “wait” or “that’s not it” — is now a baseline requirement.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"In claims processing, debt collection, appointment scheduling, and post-discharge follow-up, conversations require advancing a workflow through emotional volatility. Agents capable of handling anger, anxiety, and urgency achieve materially higher task completion rates.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Once voice enters financial services or healthcare, it becomes a production system: call recording and audit, PII and PHI handling, human escalation strategy, blocklists and sensitive keyword detection, permissions and risk controls. Fallback mechanisms are critical — how the system degrades when the model fails, audio is unclear, or emotional escalation occurs.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"}],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"link\",\"version\":1,\"rel\":\"\",\"target\":\"_blank\",\"title\":null,\"url\":\"https://substackcdn.com/image/fetch/$s_!Z7kF!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1e367eb8-9a63-4dde-ad7d-595135e2c514_2500x1862.png\"}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"type\":\"image\",\"version\":1,\"hash\":\"f7ba0988ac131e7087cebda61e2a6f45f1e2c311\",\"src\":\"https://s3-alpha-sig.figma.com/img/f7ba/0988/ac131e7087cebda61e2a6f45f1e2c311?Expires=1774224000\u0026Key-Pair-Id=APKAQ4GOSFWCW27IBOMQ\u0026Signature=nbTvs0pY7MNnAvSmsLeSTDKAbI0mov--6cHjGR4VRZHpsxWjUI~NtGPpW3Iw1qsoJIr2jHPCrCweK6i0B1uFQcJGvZBsKUFdvP3~BT5IoZjY2TUmQ-u~kLtNQfOJ1YfEJsDKS6FeZDXYL4WD3iEhRb3wzX5D2pH55BtqFTx9IvwjEEv2wQygqDGtL-D4ZOBosdlyV1jlNfnd7gSIGxkyxGPyLqZPFNwIUtQ8RZWtFzEG8aTFU6fE0aFfpNbEiTmWnKHDh3RpY3YNklszQtdvmA6t1EinCqutSaPKaUgiIyUTWzAD7loYoFdxt3WWJbJcSro8vbW2KVtZZTb5XC7Bpw__\",\"altText\":\"\",\"originalImageWidth\":800,\"originalImageHeight\":796,\"isFillWidth\":false}],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Representative companies:\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"End-to-End Speech Infrastructure: ElevenLabs, Cartesia, Sesame AI\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"ElevenLabs was founded by Piotr Dabkowski and Mati Staniszewski, initially gaining traction through high-quality TTS, and is now transitioning toward a voice agent and multimodal speech platform — targeting the enterprise voice entry layer. The company builds distribution advantages through its API and content ecosystem spanning media, localization, and AI recruiting. Its core moat is voice quality and developer adoption, not a single model capability. As OpenAI and Anthropic advance real-time voice capability, ElevenLabs’ long-term positioning looks more like a voice infrastructure layer — requiring continued expansion into orchestration and agent runtime to avoid being absorbed by foundation model capability.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Cartesia focuses on ultra-low-latency speech infrastructure. Its Sonic model series targets sub-100ms response with human-like expressiveness, positioning at the API layer rather than the application layer. The team comes from real-time systems and speech research backgrounds, with a strategy of becoming the performance baseline for real-time voice interaction — particularly suited to live customer service and robotics deployments.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Sesame AI upgrades voice from tool to always-on companion, developing voice companion devices such as Maya and Miles, pushing voice agents toward hardware and long-horizon interaction contexts. Backed by a16z and Sequoia, its core bet is on emotionally resonant voice experience and persistent context — not enterprise voice APIs. The strategy is closer to a consumer voice OS; success depends on whether it can build device and ecosystem advantages rather than a single model edge.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Vertical Voice OS: HappyRobot, Further AI, Hippocratic AI\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"HappyRobot was founded by Pablo Palafox and team, focused on a voice operating system for the logistics industry — deeply integrating voice agents with TMS and dispatch workflows to handle high-frequency tasks such as driver communication and rate confirmation. Backed by Sequoia and Accel, its advantage lies in industry-specific semantic understanding and execution loop closure, not general voice capability.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Further AI targets the insurance vertical, using voice agents to automate underwriting and claims workflows — including document follow-up requests and customer communication. The team combines insurance and technology backgrounds. At an early funding stage, the positioning leans toward workflow automation rather than foundation voice models. Its advantage is the combination of structured policy parsing with voice interaction, but it currently operates at the vertical automation layer; moat depends on depth of integration with insurance systems.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Hippocratic AI was founded by Munjal Shah and a team with healthcare and technology backgrounds, targeting a compliance-first AI nurse that uses voice to complete patient follow-up and monitoring. Backed by a16z, Kleiner Perkins, and NVIDIA, the emphasis is on safety and healthcare-grade data governance rather than general voice experience. Its advantage is vertical regulatory knowledge and healthcare process understanding; adoption depends on healthcare institutions’ trust in automated voice and compliance validation cycles.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Voice Agent Infrastructure: Retell\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Retell began as a no-code voice bot tool and has progressively repositioned as a voice agent generation platform — integrating real-time models to enable one-click generation of call logic, scripts, and CRM integrations. The team is small but iterates rapidly, operating more as a voice orchestration layer than a model company. Current advantage is stability and developer experience. The long-term risk is that as real-time capability gets natively integrated by foundation model providers, Retell’s platform value needs to extend toward workflow and enterprise integration to remain differentiated.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"root\",\"version\":1}}","itemId":"a54b6460-7b13-40d1-b52f-59d326989a10","fieldSchemaId":"3fcb25bf-f8b6-47c4-84d6-226369594160"}]}}}},"slugByItemId":{"a54b6460-7b13-40d1-b52f-59d326989a10":"openclaw-is-the-signal-our-thesis-on-long-horzion-agents","61387d11-a584-4b82-b223-7fd449542135":"the-ai-bubble-reckoning-1999-all","f650e52a-191a-4241-a8b4-12705cb62385":"ai-for-life-science-landscape","3db0c5d9-c44c-4950-a175-45c033132ebe":"ai-coding-landscape-how-agents-disrupt","c4aeef83-34d4-411e-b40c-b432dd225928":"continual-learning-next-paradigm","7b886fb4-c171-49aa-877e-475cf8b97f18":"how-openai-could-turn-the-tables","598cfcc0-365f-40c6-9f63-6631c6f5a6ff":"will-chinese-ai-leap-ahead-or-follow","91e6e6f8-7396-4490-b421-e7e2ff664063":"rl-scaling-from-research-trick-to","d285731a-bce6-45c3-9dac-26edd7829535":"pulse-how-openai-starts-outrunning","c23fe0ec-cf33-40e2-b090-b73d06099e9a":"generalist-and-the-270000-hour-advantage","80ff4031-030f-4caf-b7a9-1825fd8cb70f":"agi-2026-are-we-the-final-white-collar","940cd5e4-2e54-4b7a-941f-c0ea4e37df50":"beyond-the-cloud-are-llms-the-new","0c9ed8bc-cb92-4008-9f50-15867c8349d3":"our-thoughts-on-llm-part-one","fb158125-2004-4c70-972d-88d6846176ab":"decode-the-buzzword-why-harness-engineering-matters-now","d495e0e3-e929-4163-9d3e-0fec19ef9c2c":"beyond-deepseek-what-chinas-model"}}