The personal agent squeeze

Sixty days, five platforms

Grok Bot entered beta on August 11. Each Bot runs on a persistent cloud machine, signs into your existing tools, and works through multi-step jobs without supervision. Bots can delegate to each other, so one "chief of staff" agent can coordinate specialists. It runs on Grok 4.7 and costs roughly $120 per seat per month on Cursor Premium Teams. There is no public API or SDK yet.

Instinct opened its invite-only beta around the same time. It takes a radically different surface: no app, no chat window. You reach Instinct by texting it or calling it on the phone. Under the hood, a dedicated cloud computer runs around the clock, using vision-language models to operate websites and fill out forms the way a human would. Need to book a restaurant? Instinct Concierge places the phone call for you. Need to pay for something? It generates a one-time virtual card with a spending cap you approve. TechCrunch raised concerns about the privacy model: the terms of service grant Instinct a broad license to use "user materials," including for training.

Meta Muse launched September 8 and topped app store charts within weeks. Three tiers: free, $20/month, and $100/month. Each user gets their own virtual machine. Muse can manage tasks, break long-term goals into action plans, and interact with external services through a browser. Two weeks after launch, Amazon blocked Muse from its retail site, saying Meta hadn't authorized the agent and it didn't identify itself while browsing. That conflict revealed something about the friction ahead: when agents start acting on websites without API agreements, the web's implicit social contract breaks.

OpenAI Dots debuted at DevDay on September 29. A Dot is a persistent agent inside ChatGPT that keeps working after you close the tab. Each runs on GPT-6 Astra with its own cloud computer and browser, connecting to over 4,000 apps through the plugin ecosystem. The detail that caught my attention: when a Dot isn't actively working on a task, it does "proactive research" through read-only connections to your tools. It watches even when you don't ask, which is the listener primitive this series has been circling since day one.

Claude Duet entered research preview on October 21, the last of the five. It sends daily briefs, monitors connected apps, drafts replies, and follows up on unfinished threads. Anthropic drew a sharper privacy line than any competitor: Duet will never directly access passwords or sign-in tokens. It's available to Pro and Max subscribers in the United States during the preview.

Same architecture, five times

Look past the branding and every product ships the same stack. A dedicated cloud VM per user. Persistent state that survives across sessions. Background execution that continues after you close the app. A web browser the agent controls to interact with services that don't offer APIs.

This convergence wasn't inevitable, but it was predictable. We wrote about the three primitives early in this series: a clock (scheduled execution), a listener (event-driven wakeup), and an inbox (a place to receive and queue work). Every platform now ships all three, bundled into a single product and backed by the compute to run them continuously.

The primitives became table stakes. The differentiation moved to other layers. Distribution is the most obvious: Meta wins on reach, OpenAI on developer trust, Anthropic on safety reputation. But the trust model may matter more in the long run. Duet won't store credentials. Muse uses your data for training unless you opt out. Instinct's terms are broad enough that TechCrunch flagged them. And then there's surface area, where each product carved out a different niche: Instinct makes phone calls, Dots does proactive research when idle, Muse controls macOS apps.

We've tracked the cost of always-on compute and the shift from foreground to background in earlier posts. What changed in this quarter is that the economics went from theoretical to real. Five companies are now paying for a VM per user, continuously, at scale. The viability question is settled. What remains is who can absorb the cost long enough to keep the lights on.

Where the middle went

The space bifurcated. On one side sit platforms with billions in compute budget, subsidizing personal agents as a strategy to deepen lock-in. Meta offers Muse for free because every interaction feeds data back into the ad-targeting pipeline. OpenAI bundles Dots into ChatGPT Pro because persistent agents increase engagement and API stickiness. Anthropic ships Duet as a research preview for existing subscribers, building toward a consumer product that justifies the model training investment. Google shipped Gemini Spark at I/O earlier in the year with the same playbook.

On the other side, Instinct reportedly raised at a $10 billion valuation. That number makes sense only if you accept that offering an always-on cloud computer to every user, without an existing platform to amortize against, requires enormous capital. Instinct has no model to sell and no social network generating ad revenue, just a product that people text, where every message triggers compute on a dedicated machine. At that burn rate, the capital itself becomes the competitive advantage.

The middle, where small teams built proactive agents on top of general-purpose infrastructure, got squeezed from both directions. Val Town is the most visible casualty, but the pattern extends further. When platforms hand out VMs at cost or below cost, building agent infrastructure as a standalone business gets very hard. The infrastructure layer that indie builders relied on is disappearing, and the replacement is platform-specific: Grok Bot runs on Cursor, Dots runs inside ChatGPT, Duet runs inside Claude, Muse runs inside Meta's ecosystem. The agent's compute is inseparable from the platform's identity.