Claude Duet: Waitlist, Usage Limits and Approvals

Anthropic introduced Claude Duet on October 21, 2026 as a proactive assistant that can review connected work apps, identify items that need attention and continue assigned tasks while the user is away. It is a research preview, not a general release: Anthropic is initially inviting US Claude Pro and Max subscribers through a waitlist. The most useful distinction is between what Duet can prepare independently and what still requires explicit human approval.

What Claude Duet does differently

Duet combines a morning brief with periodic checks of connected accounts and a persistent conversation that accepts typed or spoken requests. Anthropic says it carries over Claude memory, skills and connected apps. A hosted virtual machine and browser allow work to continue even when the user's laptop or phone is switched off. Tasks described by Anthropic range from scheduling and drafting sales follow-ups to assembling dashboards, financial forecasts, prototypes and web apps. These are announced capabilities and early-user examples, not independent benchmarks of accuracy or completion rates.

Background checks are not continuous monitoring

Duet checks connected accounts every few hours rather than watching them continuously. Anthropic says this trade-off helps preserve plan usage limits, and users can tell the assistant which updates matter and how frequently they want to hear about them. That makes it a candidate for periodic inbox or workflow triage, not a substitute for a time-critical incident alert or guaranteed instant response. For operational alerts, teams should keep their existing event-driven monitoring and escalation paths.

Where human approval still applies

Anthropic says Duet can read, draft and prepare work autonomously, but asks before sending a message, making a payment, signing a document or deleting something. It also prohibits certain activities even with approval, including securities trading and moving money between accounts. The company describes credential isolation, automated action screening and testing against prompt injection. Those protections do not eliminate the need to review connected-app permissions or check the contents of a proposed action before approving it.

What to test before trusting it with a workflow

1. Choose a low-risk connected-app workflow, such as summarizing follow-ups without sending messages.
2. Confirm that the necessary connector is available to your account and limit its permissions to the minimum required.
3. Specify which events deserve notifications, then observe whether periodic checks surface them at a useful time.
4. Review every proposed external message, purchase or destructive change before approval.
5. Measure missed items, incorrect priorities, usage consumption and human review time before expanding access.