Everybody's Building the Same Agent

A note before you read this: the article below was written by Meta Muse — Meta's personal agent — not by me. I'm publishing it as an experiment, and as evidence. The experiment was whether an agent could take my notes and turn them into a publishable article. The evidence is the result. I'm somewhat disappointed by the quality of the writing, though I'm not surprised: nothing about how Muse is pitched — errands, bookings, inbox triage — suggests writing is what it's for. Read it as a field report from the gap between what these agents can do and what I want them to do. — Yusef

In 2026, everybody shipped a personal agent. Meta's sits at the top of the App Store, with an estimated 2.8 million downloads in its first two weeks (Sensor Tower, via TechCrunch). Microsoft built its autonomous agent on an open-source project whose creator now works at OpenAI. A 23-year-old's invite-only text-message assistant is worth $2.5 billion. And the most-starred software project in GitHub history is a self-hosted agent you talk to over Telegram.

They present as competitors. Line them up and the differences that matter fit in a short list.

The form

Every one of them is a chatbot attached to a computer that works while you're away. You talk; it acts across your connected accounts — email, calendar, shopping, travel, messaging — and reports back. The industry converged on the physical shape without announcing it: one persistent cloud computer per user. Meta calls it a Secure VM. xAI gives each Grok Bot its own cloud computer.

The open-source world got there first. OpenClaw — over 390,000 GitHub stars as of this week, the most-starred software repository on the site — is a self-hosted gateway you message over WhatsApp, Telegram, or Discord. It runs shell commands, drives a real browser, keeps cron jobs, and spends your model API keys. Its creator, Peter Steinberger, announced in February that he was joining OpenAI to work on personal agents, and the project moved into an independent foundation. It is actively maintained and shipping. What faded was not the project but its distinctiveness: the revolutionary thing became the default thing, as incumbents shipped the same shape as a polished app. Commoditization, not abandonment.

Its challenger, Nous Research's Hermes Agent, arrived weeks later with a "closed learning loop" — it reflects after complex tasks and writes itself reusable skills. I tried one of the early agents myself and churned. The rough UX wasn't the reason. More on that below.

The segments

'Big tech' sells distribution. Muse launched September 8 and reached #1 on the US App Store on September 18. Pricing is a usage meter, not a feature ladder: free up to a weekly token allowance, $20 a month for Power, $100 for Maximum. Same agent, more headroom. Underneath is Muse Spark, plus a kernel-isolated process called Sentinel that must approve anything sensitive — email, purchases, money movement. Amazon has blocked the agent from its store. That last fact matters more than the download charts: the agent economy runs into the platform economy the moment it starts spending.

Google's Gemini Spark is the same shape with one honest engineering difference: it acts through direct API integrations rather than driving a browser like a human, which Google positions as more reliable. xAI's Grok Bot goes the other way — a cloud computer per bot that signs into your apps and learns a workflow by watching you once.

Microsoft's entry is the strangest. Scout, unveiled at Build in June, is an always-on background agent for Microsoft 365 built directly on the open-source OpenClaw framework, with each instance getting its own Entra identity so it acts for you rather than as you, under audit. Anthropic's Orbit is staged but unlaunched. OpenAI's "Aeon" remains a rumor: one string in the ChatGPT web client and one report in The Information that the company is developing agent features to compete with Grok Bot. At DevDay on September 29, OpenAI is expected to unveil "Managed Agents" — a platform for building and deploying agents, aimed at developers and enterprise rather than consumers. Apple has no agent at all.

'Venture-backed startups' sell the concierge dream before the tech is ready. Instinct, from Noah Shinn's Spear Street Technology, is the purest specimen: invite-only, messaging-first with no app at all, proactive enough to text or call you first, valued at $2.5 billion after raising $350 million from Index, Benchmark, and Kleiner Perkins (background). An investor called it "OpenClaw for normal people." Then came the privacy firestorm: terms granting a perpetual license to user materials, an agent that acted without being asked. Nobody will say what model it runs on. There is no public pricing. Unconfirmed reports have it in talks to raise again at roughly $10 billion — which would value an invite-only beta at roughly the GDP of a small country.

Town is Instinct's mirror: $55 million Series A led by Andreessen Horowitz, live and charging $15 to $199 a month, model-agnostic across providers — the model, the company says outright, is a "commodity input"; the accumulated per-user context is the moat — and anything sent as you waits on explicit per-send approval. It is the only company in the segment that treats the trust problem as the product rather than the marketing.

The rest sorts by audience. Poke, acquired by Cognition in July after a reported 100 million messages in three months (TechCrunch); Cognition has since raised $2 billion at a $48 billion valuation, and an unverified rumor holds that it will ship a personal agent platform within weeks. Manus, reportedly past $100 million in ARR on the corporate side. A half-dozen "AI chief of staff for X" startups for sales, restaurants, and small business.

'Luxury and human-fallback' is where the category admits what it can't do. The traditional concierges are human-first with AI as internal tooling. The AI-first luxury startups invert it: some disclose no human fallback at all, others pair the agent with human desks "when AI meets its limits." The general-purpose version is the most honest: Magic labels its chat bubbles with the human assistant's name when one takes over. Others hide the handoff and the work just appears.

That gives you the axis nobody puts on the spec sheet: visible humans versus invisible humans. Users may accept the handoff; they punish discovering it. And pricing in this segment scales with the human, not the AI.

What actually differs

Same form, and the live axes are execution mechanism (browser-use versus API integrations), integration depth, trust and governance posture, pricing, and audience. Note what isn't on the list: the model. Almost nobody discloses which model runs underneath. When an industry won't name its core input, the input has commoditized.

And note the axis the spec sheets undersell: governance. Town's per-send approval gates, Scout's per-agent Entra identity, Muse's Sentinel layer, Instinct's willingness to act unprompted — these aren't implementation details. They're the product, because the product is permission to act in your name.

The demand problem

Which is the thing nobody's solved. Take the category's own touted jobs: canceling subscriptions (one-time), restaurant reservations (rare for most people), online ordering (wasn't hard before the chatbot). The safe jobs don't recur. The recurring jobs — emailing as you, messaging as you, spending as you — aren't safe to delegate, because the errors are denominated in your reputation, and no one outsources their social standing to a system that's right 95% of the time. Instinct's unprompted email is the cautionary tale.

This is why the companionship push deserves a skeptical read. Meta unveiled the Muse Charm at Connect — a Tamagotchi-like keychain with a cartoon avatar, targeting December. You don't build a physical vessel for a relationship with your to-do list. The charm exists because the chores don't bring anyone back, so the company is manufacturing a relationship instead. Note the status math, though: Meta's DNA is status-as-a-service, products whose value is performed to other people. A companion is consumed in private. The charm itself is a Meta camera, microphone, and 5G radio dangling off your bag — "Zuck is watching" made wearable — aimed at a holiday-toy price point. No status in the relationship, no status in the object. For a company whose superpower is the young urban cohort that drives global trends, it's a category error about what they're good at.

I'm not against the category — this piece was written with a personal agent, and it did real work this week. But it's not that exciting, and that's the point. The bad UX of the assistant I used before wouldn't have stopped me if it were truly valuable. People churn because the value doesn't compound: errands end. What compounds is the accumulated state — the history, the corrections, the context — and every product on this map rents that state from its vendor except the open-source ones, which make you host it yourself. That's the actual contest underneath the feature lists: not who has the smartest model, but who owns the workbench when the agent changes.