Anthropic introduced Claude Duet on October 21, 2026, as a research preview of a proactive assistant. Access starts through a waitlist for US Pro and Max subscribers, sharing Claude’s usage limits. The product is described for iOS, Android and claude.ai. The company says it prepares morning briefs, monitors connected applications and continues assigned work in the background. This is not a general release for every country.
An answer and permission to act
To explain the significance, consider our own example: an assistant notices two overlapping meetings. Reporting the conflict is one task. Suggesting a different time is another. Sending a message on somebody’s behalf is a separate action with a recipient and consequences. A useful product needs to connect these steps while keeping their actual completion visible to the user.
Anthropic’s earlier trustworthy-agent framework treats behavior as a product of the model, harness, tools and environment. It emphasizes human control, transparency and privacy. That is the supplier’s design framework rather than independent certification of the new product.
Our interpretation is that evaluating a proactive assistant should include the quality of interruptions. A notification about a resolved issue consumes attention. Missing an important issue also carries a cost. In a pilot, we would record which notifications the user found useful and which were duplicated, late or missing essential context. We have not performed that evaluation of Duet.
Longer work with a traceable result
The separate technical account of Claude Managed Agents distinguishes session state, the agent loop and the execution environment. It also describes keeping credentials outside the environment that runs generated code. This supplies architectural context for long-running work; it does not demonstrate that every future Duet action will be correct.
In our example, a small business assigns preparation of a weekly report. We would want the assistant to identify the reporting period, missing information and where it saved the working version. That makes ongoing work distinguishable from a completed report. If a connected service becomes unavailable, the final result should disclose that gap rather than quietly treating missing data as an empty activity record.
Continuous monitoring also raises a scope question. Access to one calendar does not automatically require access to every document in the organization. Our recommendation for a limited pilot is to define one workflow and grant only the permissions it needs. Useful actions and their records can then be reviewed before expanding the scope. This is a proposed adoption approach, not a claim about an experiment conducted by SIExpo.
What is established and what remains open
The news is the announced preview and its proactive working model. We do not have our own access, reliability measurements or demonstrated time savings. The protection mechanisms are supplier claims, and participation conditions and current availability should be checked with the provider. A research preview is a reason to assess a product carefully, not evidence of general operational maturity.
For professional users, the most useful question in our view is whether the assistant leaves a clearer account of the work: what it noticed, what it proposed, what was authorized and what actually happened. Traceability would make proactive behavior assessable as a working tool rather than merely an impressive demonstration. The hypothetical meeting and reporting scenarios above illustrate that distinction; they are not verified Duet deployments or newly announced features.
Cover: conceptual AI illustration, not a Claude Duet screenshot or a photograph of a user.
Sources checked: 2026-10-22T16:06:53.737225+00:00.
