.jpg){kind=link}
Anthropic launches Claude Duet: an AI agent that works in the background — against your usage limits
Anthropic launched Claude Duet on October 21, 2026, a proactive, "always on" agent in research preview for Pro and Max subscribers in the US. It keeps running after the app is closed, can check invoices in Stripe and make purchases with your approval — but all that background work draws from the same usage limits as chat, which Anthropic itself warns may drain them "faster than with chat alone."
Anthropic launched Claude Duet on October 21, 2026, a proactive AI agent now rolling out in research preview to Pro and Max subscribers in the US, with access granted gradually via a waitlist (9to5Mac). The agent is described as a personal assistant or "chief of staff": you send it a message, written or spoken, the way you would message a person, and it works independently in the apps you have connected — including after you have closed the app (Cryptobriefing).
All information about the launch comes from two secondary outlets, 9to5Mac and Cryptobriefing, both of which reported the same day and agree on the core facts. No primary documentation from Anthropic — press releases, blog posts or support pages — underpins this story. That means details such as model choice and safety features are the company's own claims, not independently verified.
How it works
Duet is not a new chat tab, but an agent on its own cloud-based machine. It therefore keeps working even after the user has closed the app, unlike a regular chat that only responds when you type. According to 9to5Mac, Duet reuses the same memories, skills and connections as regular Claude, so a user who has already connected Claude to calendar or email does not need to rebuild the context.
Anthropic says Duet is powered by Claude Opus 5.5. The claim is provided by the company and relayed by both outlets, but it has not been independently technically verified.
The interaction is built around proactivity: the agent monitors connected apps such as email and calendar and carries out tasks like scheduling and invoicing without the user having to approve each step, Cryptobriefing writes. When something requires human input, the agent says so — and waits.
Integrations and use cases
Three concrete integrations are given at launch:
- Stripe: The agent can check for outstanding invoices and draft payment reminders.
- Link and Shop Pay: Duet can make purchases, but only with the user's approval.
- Its own cloud-based browser: For sites without a connection, the agent can use a browser running in the cloud.
Cryptobriefing's analysis places the target audience among professionals and small business owners, with scheduling and invoicing as the kind of administrative grind the agent is meant to take over. The outlet also characterizes the launch as a move in a field where OpenAI and Meta are also moving into personal agent technology — but that is the outlet's characterization, not something Anthropic itself is quoted on.
The safety model: designed to ask
This is the limiting element that separates Duet from a freely running agent. Anthropic describes the product as "designed to ask before it sends a message, makes a payment, signs a document, or deletes anything" (9to5Mac). In addition, only the user themselves can mark tasks as completed — the agent cannot declare a task finished on its own.
Worth noting: this is the manufacturer's own description. Both the "designed to" wording and the claim that "only the user" can complete tasks are relayed from Anthropic's launch material; no independent testing of how reliable these approval gates are in practice appears in the source material. For a product that can send messages and make payments, whether the gates hold up in everyday use is precisely the central open question.
The price in practice: the same usage limits
Duet is included at no extra cost for Pro and Max subscribers — but all background activity draws from the same usage limits that apply to chat and other Claude features (9to5Mac; Cryptobriefing). Anthropic itself notes that this may lead users to hit the limits "faster than with chat alone."
That is an acknowledgment with practical significance: an agent that monitors apps and works continuously in the background consumes resources, by definition, while the user is not actively asking for anything. How quickly the limits actually drain for a typical subscriber is impossible to say from available sources — no figures or data on consumption rates have been published. But the fact that the company itself flags the risk suggests it is real enough to mention in the launch announcement.
Who gets access
Access is deliberately gradual:
- Who: Only Pro and Max subscribers. Team and Enterprise plans are not covered yet; Anthropic says expansion to these plans is planned, but gives no timeline (9to5Mac).
- Where: US only at launch.
- How: Invitations go out in waves via a waitlist. The exact waitlist pace or how many people get access is not stated.
- Where you use it: Dedicated apps for iOS and Android, plus claude.ai on the web, for invited users.
The naming story — with caveats
Cryptobriefing reports that the project was internally known as Claude Loop, and that Anthropic switched to Duet in connection with the mobile rollout. The same source ties the launch to a change in September 2026, when Claude Cowork was folded into the main chat interface. Both of these details appear only in Cryptobriefing and are not confirmed by other sources or primary documentation — they should be read as reporting on undocumented internal matters, not as established facts.
What remains to be answered
Several key questions remain open after the launch:
- Real-world usage consumption. Anthropic acknowledges that background work draws on the limits, but no data shows how fast. For an "always on" agent, this may be the single most important unknown for whether the value holds over time.
- Guardrails under pressure. The approval gates before messages, payments, signing and deletion are design principles, not tested results. Independent evaluations are so far lacking.
- Rollout. The waitlist pace is unspecified, Team and Enterprise users have no timeline, and it is unclear when or whether access will expand beyond the US.
Claude Duet is thus a concrete step toward AI agents that work independently in connected apps — but with a pricing setup and a safety model that, so far, have only been described by the party selling them.