Anthropic introduced Claude Duet, a research preview of a personal agent intended to help between explicit conversations. Rather than waiting for every individual prompt, the product can prepare briefings, check for developments, and carry out background work in a virtual environment.
The announcement describes access through mobile applications and the web, beginning with a waitlist. Duet can read, draft, and prepare actions, while seeking approval for consequential steps such as sending messages, making payments, signing, or deleting. Background activity draws on subscription usage. That makes the frequency and usefulness of checks part of the product's practical value.
Proactivity needs an interruption budget
An assistant that notices everything can quickly become another source of work. A useful starting point is one recurring responsibility with an observable result: prepare tomorrow's meeting materials, identify a changed travel detail, or flag a document that requires a response.
Each workflow should distinguish between a routine update and something worth interrupting a person for. A daily digest might be sufficient for a stable project. A time-sensitive conflict may justify an immediate alert. Treating both events identically would waste attention even if the underlying summaries were accurate.
The same discipline applies to background spending. A user should be able to explain what a recurring check is expected to discover and when it should stop. Repeatedly inspecting an unchanged source has little value simply because the agent is capable of doing it.
Approval should expose the actual decision
A good approval request presents the destination, the relevant content, and the consequence of proceeding. It should also make cancellation easy. A vague invitation to approve an entire ongoing activity gives a user less useful information than a specific proposed action.
Personal context will change over time. A preference captured during one project may be inappropriate during the next. A trial should therefore include a deliberate correction: change a standing instruction and observe whether future preparations respect the new boundary.
The most informative success measure is saved attention after review. Count useful preparations, unnecessary notifications, corrections, and time spent supervising. That gives a clearer picture than counting tasks initiated in the background.
Duet makes proactive assistance a concrete product direction. Its promise will be realized when a user can delegate a small responsibility, understand what happened, and comfortably leave the assistant quiet until it has something useful to contribute.
Have a correction or a useful perspective?
Get in touch with the publication.


