The interesting part of Claude Duet is the interruption. A proactive assistant has to decide when to leave you alone, when to bring something up, and when an action needs your approval. Getting the answer right is only part of that job. 💡
Anthropic opened an early research preview on October 22. Its announcement describes an assistant that surfaces work, continues tasks while you are away, and checks before acting on your behalf. Access starts with a US waitlist for Pro and Max customers. This is a look at the announced design, rather than a hands-on review. [1][2]
My illustration of an action boundary. This does not reproduce Duet's internal architecture.
A useful suggestion can still arrive at the wrong time
Imagine a calendar assistant notices that a meeting overlaps with another appointment. It can surface the conflict, prepare a proposed change, or actually move the meeting. Those steps have different consequences.
The first consumes attention. The second creates something you can inspect. The last changes other people's calendars.
I would judge a proactive assistant on whether it keeps those consequences visible. A polished notification that hides an unexpected calendar change would make the experience worse, even if the new time were perfectly reasonable.
That is a design question for any product built around agents: what event should bring the user back into the loop? 🧭
Memory helps with relevance; authority still needs a boundary
Anthropic says Duet carries over memory, skills and connected apps from a Claude account, and is built around Opus 5.5. It also describes a classifier intended to guard sensitive information and says the assistant does not directly see passwords. Those are vendor descriptions, not results I independently verified. [3][4]
Remembering a preference can make a suggestion better. It does not automatically answer whether a particular message should be sent or a particular file should be shared.
For example, “I prefer afternoon meetings” gives an assistant context for a proposal. It leaves open which meeting, who receives an update, and whether the proposed change is still appropriate now.
⚠️ The common mistake is to treat remembered preferences as an open-ended instruction to act. A useful product can make the intended action concrete: show the recipient, the change and the current reason for it. Then the user has something meaningful to approve.
What I would look for in a preview
My first question would be whether suggested work is easy to dismiss or correct. An assistant needs a way to learn that an apparently urgent item can wait without turning every correction into another task.
I would also look for an activity history that connects an action to the instruction that authorized it. “Done” is satisfying; knowing what changed is more useful when something goes wrong. 🔍
Finally, I would check how the product behaves when it is unsure. Preparing a draft can be valuable progress. Repeatedly asking a vague question can create more work than it removes.
Anthropic explicitly says the preview can misjudge urgency. That is the limitation I would keep closest to the launch promise. A proactive assistant earns its place by making its interruptions useful. [2]
References
- Claude's official Duet announcement, October 22.
- Preview availability and urgency limitation.
- Model, memory, skills and connected apps.
- Announced safety approach.
Asset credit: original explanatory diagram by Shivam Pandit. Illustrative, not a product screenshot.
AI disclosure
AI was used to draft the article.
Top comments (0)