By ByteBot
Anthropic launched Claude Duet on October 21 in research preview — a standalone app where Claude monitors your connected accounts, sends a morning brief, and keeps executing tasks after you close your laptop. It runs on Opus 5.5 inside a cloud VM, reaches 6,000+ app integrations, and hard-stops before it sends any message, makes any payment, or signs any document on your behalf. That last part is not a footnote — it is the entire argument Anthropic is making about what a background agent should be.
What Duet Does, Concretely
Each morning, Duet scans connected apps — Gmail, Slack, QuickBooks, HubSpot, Gusto, Docusign, Shopify, and hundreds more — and surfaces the items that need attention. During the day it checks every few hours and pings you only when a decision is required. The rest of the time it keeps working in its own cloud VM, with its own browser, independent of whether your device is on.
In research preview, six task categories are functional today: drafting unpaid invoice reminders from QuickBooks or Stripe data, preparing Gusto payroll runs with hours and bonuses pre-filled, building HubSpot follow-up emails from deal history, drafting LinkedIn job postings, preparing Docusign agreements for your signature, and lighter development work including financial forecasting and prototyping. The 6,000-connector claim draws from the same Claude Marketplace already available in Claude.ai.
Three Labs, 22 Days, One New Category
OpenAI launched Dots on September 29. Meta launched Muse on September 30. Anthropic launched Duet on October 21. Three frontier labs shipped background agents in 22 days. This is not an experiment — this is a product category crystallizing in real time.
The three are not interchangeable. Dots wins on integration depth: native Slack and Teams support, a direct pipeline into Codex, and a programmable Agents API. Muse wins on raw autonomy: it can browse, fill forms, and place orders without pausing for approval on reversible actions. Duet wins on trust architecture and model quality: Opus 5.5 is the strongest reasoning model among the three by most benchmarks, and the approval design prevents the agent from taking actions you will regret.
| Feature | OpenAI Dots | Meta Muse | Anthropic Duet |
|---|---|---|---|
| Launched | Sept 29, 2026 | Sept 30, 2026 | Oct 21, 2026 |
| Model | GPT-6 Astra | Muse proprietary | Claude Opus 5.5 |
| Autonomous browsing | Read-only proactive | Full (reversible acts) | VM browser, gate-gated |
| Slack/Teams native | Yes | No | Via connector |
| Public API | Agents API (Codex) | Not announced | Not yet |
| Approval model | Soft (suggests) | Autonomous (reversible) | Hard gate required |
Which one fits your workflow depends on what you are delegating. For a coding copilot that watches your repo and acts inside Slack: Dots. For an agent that books, shops, and browses on your behalf: Muse. For an agent that handles financial and HR administration without unilateral action: Duet.
The Approval Gate Is the Point
Before Duet sends any message, makes any payment, or signs any document, it stops and shows you what it is about to do. This gate is enforced at the infrastructure level by the same auto-mode classifier that governs Claude Code — it cannot be toggled off through a prompt or a settings switch. A second tier of actions — trading securities, moving money between accounts — is refused even with explicit user approval.
That will frustrate users who want hands-free operation. But consider the alternative: an agent with access to your Stripe account, Gusto payroll, Docusign, and email that acts first and logs it afterward. The approval gate is not a limitation Anthropic plans to remove when the preview ends. It is the product stance: AI should prepare and propose, not decide and execute.
An agent that never sends a message without your sign-off sounds like half an agent — until you consider it has access to your payroll system.
What Is Still Missing
Duet has no public API, so programmatic triggering of its workflows is not possible today. It is US-only on a waitlist for Pro and Max subscribers. There is no team mode or shared-agent workspace announced. Android support is coming; at launch it is iOS and web only.
The more meaningful open question is whether the approval gate can scale gracefully. An agent that pings you twelve times a day for confirmations is not a productivity gain — it is a new inbox. Anthropic will need to either raise the autonomy floor or invest in batching approvals into a single daily review queue. The New Stack’s breakdown of Dots illustrates how much smoother a softer approval model feels in practice. That pressure will not go away.
The Bet
Anthropic is wagering that users will ultimately trust the agent that asks permission more than the one that does not, and that trust translates to stickier usage and fewer catastrophic failure stories. That may be right. It may also be that users default to whichever agent does the most with the least friction. The next six months will settle it. If you are a Pro or Max subscriber in the US, the waitlist is open at claude.ai.
