Claude Duet entered a waitlisted research preview on October 21, 2026, initially for US Pro and Max subscribers. Anthropic describes a proactive app that notices work across connected services and continues tasks in the cloud. A second official article, published October 23, explains its approval and information-handling controls. Together, the announcements describe a product direction with meaningful access limits.
Original editorial illustration; this is not a product screenshot or a measured result.
This report is current to October 27. It analyzes the announcements and their implications rather than claiming access, reporting a deployment trial, or treating a preview as a general release. The distinction matters because the product's appeal depends on both what it can prepare and where it must stop.
Claude Duet is a preview with a specific audience
The October 21 announcement identifies professionals and small-business owners as the intended audience. It describes a morning brief, periodic checks of connected apps, and background work that does not require the user's device to remain on. Preview activity shares the subscriber's existing Claude usage limits.
Our interpretation is that the product shifts the starting point of a task. In a conventional conversation, the user notices the work and asks for help. A proactive service also participates in deciding which unfinished work should reach the user's attention. That is a larger product responsibility than generating a response after a prompt.
The value of that shift remains an empirical question. A useful alert could prevent a missed commitment; an unnecessary alert could add another interruption. These are possible consequences of the design, not outcomes measured for this article. The announcement alone cannot establish how well that judgment will work for a particular business.
Availability has several different meanings
The launch describes access through a waitlist, with standalone mobile apps and a web presence. It also identifies a richer Shopify visual editor and 1Password sign-in as forthcoming. Those details should not be combined into a claim that every advertised workflow is already available to every subscriber.
| Status in official material | Reading appropriate for this report |
|---|---|
| Research preview by waitlist | Announced access for an initial audience; admission is not guaranteed |
| Described preview functionality | A company statement about the early product, not an independent test |
| Coming-soon functionality | A roadmap statement, with completion unverified here |
This table describes evidence categories, not a feature scorecard. A buyer can find the direction promising while still needing to establish eligibility and actual account access. A report that removes those distinctions would make the announcement sound more complete than the sources justify.
There is no general-release date asserted in this article. Neither inclusion in a subscription nor the existence of an app listing should be taken as proof that a reader has entered the preview. The correct current status remains the one the official announcement specifies.
Human approval is part of the product architecture
Anthropic's October 23 safety article describes checks separate from the working model, user approval for important actions, and limits on handling sensitive information. It also acknowledges that safeguards can miss a problem or flag a harmless message. These are provider descriptions and caveats, not an assurance from this publication.
Preparing work and committing it create different responsibilities. A draft can be reviewed before it affects another person. An external action has a destination, an audience, and a consequence. Preserving that boundary is central to understanding the proposed service, rather than an incidental warning attached to the launch.
Approval also creates a design problem worth watching. A review request must give the user enough context to judge the action without recreating the entire task. Too little information makes consent superficial; too much can make every task feel like another project. This report identifies that tension without claiming a result from using Duet.
Disconnecting access and forgetting context are distinct
The safety article says disconnecting an app stops new access, while information already learned persists until memory is reset. It also describes isolation of credentials from the model. These points concern different boundaries: future access, retained context, and authentication material should not be treated as interchangeable concepts.
For a team, the practical implication is clarity about what an assistant knows and what it may still retrieve. Removing a connection can be a sensible control without erasing everything incorporated into prior work. An operating process needs language precise enough to explain those distinctions to the person responsible for the account.
This is analysis of the documented design. It is not a privacy certification, a security audit, or evidence that a particular implementation was tested successfully. Readers evaluating sensitive work need the current official documentation and their own requirements, rather than an assumption that one control solves every information-handling concern.
The next evidence should concern ordinary work
The most useful follow-up to a proactive-agent launch is evidence about routine tasks: whether the relevant item was found, whether its urgency was understood, and whether the final handoff preserved the user's intent. A dramatic demonstration can show possibility without answering those more ordinary questions.
Usage accounting belongs in that assessment as well. Background work consumes resources even when the user is not watching a chat window. A business deciding whether the service fits its day should be able to distinguish an informative notification from an expensive or unnecessary sequence of work.
The site's productivity category provides context for related tools. It does not turn one vendor's preview into proof that proactive agents have solved administrative work. For now, the news is a bounded release and an articulated control model, with customer outcomes still requiring evidence.
FAQ
Is Claude Duet generally available to every Claude user?
The October 21 source describes an initial waitlisted research preview for US Pro and Max subscribers. This report does not establish broader access or a subsequent general release.
Were the forthcoming integrations tested for this report?
No. Roadmap items retain their announced status. No access trial, independent performance measurement, or completed rollout is asserted.
Why distinguish disconnection from a memory reset?
The October 23 documentation describes them as different controls. Stopping new access should not be described as deleting everything the agent already learned.
Conclusion
Claude Duet is a concrete October preview announcement with a clear ambition: combine connected context, background work, and a user who remains responsible for consequential actions. The meaningful next question is how that arrangement behaves in ordinary use. Waitlisted access and forthcoming capabilities should remain visible in any account of what has launched.