Anthropic introduced Claude Duet’s proactive-agent research preview on October 21, 2026, followed by a safety, privacy, and security explanation on October 23. The hosted application is designed to work across connected services in the background. Its approval gates and account controls are central to deciding which tasks it should handle.
Claude Duet is a hosted application preview, not an announced Duet API or downloadable-weight release. Anthropic’s announcement describes an agent built around Claude Opus 5.5. Neither that announcement nor the safety explanation announces a Duet developer endpoint or model-weight download.
What launched—and who can access it
Anthropic’s October 21 Help Center release-note entry records the preview for Pro and Max subscribers. As checked on November 1, the current product page lists US Pro, Max and Ultra subscribers, still through a waitlist; Team and Enterprise plans are not included. That distinction matters: today’s eligibility wording should not be treated as proof that every listed plan was part of the original launch.
The product is offered through standalone iOS and Android apps and the web. Its product FAQ says background checks share the subscriber’s existing Claude usage allowance. Inclusion in a subscription therefore does not mean unlimited unattended work—or immediate access after joining the queue.
Proactive does not mean real-time or unrestricted
The October 21 announcement describes morning briefs and checks of connected accounts every few hours. Duet works on a cloud computer, so tasks can continue without the user’s device remaining on. Examples include finding unpaid QuickBooks invoices and drafting reminders, or preparing HubSpot follow-ups.
For an owner juggling receivables and sales, our assessment is that the useful change is earlier preparation: a reminder can be ready when the owner next checks in. It should not be interpreted as a promise of instant detection. A workflow needing a response within minutes would require separate verification of its timing rather than relying on the phrase “proactive agent.”
Anthropic also acknowledges that Duet can misread conversations or misjudge urgency, and the announcement says users cannot yet see its browser. Those limitations make it a candidate for supervised trials, not evidence that a business process can run without oversight.
The October 23 explanation adds important boundaries
Anthropic describes a separate action check that evaluates whether an operation matches the user’s intent. Approval screens expose the action, destination and content. Securities trading and moving money between bank accounts remain prohibited even with approval; some consequential operations cannot receive a standing “don’t ask again” exemption.
A separate sensitive-information classifier checks outgoing messages. Credentials, full card and bank-account numbers, and government identifiers fall into a never-share tier. Other private information requires consent, although an explicit instruction can itself supply that consent. Anthropic warns that these checks can miss sensitive content or flag harmless material.
There is also a less obvious retention boundary: disconnecting an app prevents new access, but information already learned persists until memory is reset. Credential isolation is therefore not the same thing as erasing business context.
What the preview means for a business pilot
Our operational reading is to evaluate Duet on the quality of decisions it prepares, not the volume of activity it generates. For a receivables pilot, that means checking whether a proposed reminder names the correct customer, invoice and recipient before approving it—not assuming that a fluent draft is accurate.
Three questions follow directly from the preview’s design:
- Who reviews pending actions? Assign a reviewer so work does not stall behind an unattended approval request.
- Which context is appropriate? Start with the minimum accounts needed, and decide when retained memory should be cleared.
- What constitutes success? Track useful drafts, corrected mistakes and missed priorities alongside the usage consumed.
The next evidence worth watching is whether Anthropic changes preview eligibility, approval calibration or task visibility. Until then, the October 21 and October 23 releases support a narrower conclusion: Duet offers proactive preparation with layered controls, not unrestricted autonomy or a newly released deployment platform.