Skip to content
kenjtecJournal
Search the Journal

Claude Duet: the agent that keeps working even between conversations

Anthropic’s research preview of 21 October 2026 moves work into the background. What matters is which state an agent tracks and when it really needs you.

Tobias takes a card from a washing line of cards between two conversations. AI illustration.

A delivery date changes between two conversations. A classic chat only learns about it when you bring the new information along. A proactive agent, by contrast, is meant to keep watch, carry its work forward and bring you back in when something relevant changes. That is the interesting idea behind Claude Duet.

Anthropic announced Duet on 21 October 2026 as a research preview. Access starts via a waiting list for US subscribers to Pro and Max. The vendor’s description mentions background work with connected apps and approvals before important external actions. Background work shares the usage limits with the rest of Claude. This article is not based on access of our own or on a product test.

From a single task to ongoing responsibility

With a single question, the end is usually clear: the answer is there. Ongoing responsibility, by contrast, needs a state that the agent interprets sensibly across several observations. Which information is confirmed? What is still open? Which new detail changes the next decision?

Our own example is a materials order for an exhibition. What is relevant are the delivery date, the set-up time and a room approval that is still pending. “Take care of the exhibition” leaves almost everything open. “Report a delivery date that falls after the confirmed set-up and name the source”, on the other hand, describes a specific responsibility.

A change is not yet a reason to report

A delivery page can reword a heading without the date changing. Conversely, a single date can have an important consequence. A useful report must therefore explain the context: what has changed, compared with which known state, and for which decision is it relevant?

Set up a small monitoring contract. It can contain the explicitly permitted sources, the confirmed reference state and a reporting trigger. In the example, a missing source also counts as a state of its own. “Cannot be determined” means something different from “still on time”.

Monitor only the named exhibition sources.
Reference: set-up on the confirmed date, delivery status as per the attached note.
Report a new delivery commitment that falls after the set-up, or a withdrawn room approval.
State the change, the source location and the decision affected.
Prepare possible follow-up questions; do not send them.

Keeping preparation and external effect apart

Drafting a follow-up question and sending it to a supplier are different actions. In a draft, a person can correct an unfortunate wording. A message that has been sent, on the other hand, can look like a binding commitment. The task contract should describe this boundary before the background work begins.

The announced product approvals are another layer. They do not replace the decision about which task was delegated in the first place. Conversely, an approval interface described by the vendor does not yet prove fully tested protection for every conceivable workflow. That would require an actual test of the specific action.

After a confirmed change, the reference state also needs a deliberate update. Otherwise the same, already known date can appear again as a new report at the next comparison. Who maintains the working state is therefore part of the task contract.

The first sensible area stays small

A manageable area of responsibility makes misbehaviour visible. Superfluous reports are as interesting as missed changes. If every run merely rephrases known information, the agent mainly produces extra reading. A helpful update should answer a new question or prepare a decision.

Anyone who gets access can start with artificial source states: unchanged, a relevant change of date and an unreachable source. Watch which information actually appears in a report. These are our own suggestions for getting started, not reported Duet results.

The news value lies in the shift from conversation to supervised responsibility. How far this concept goes depends on access, sources, usage limits and understandable assignments. The research preview and the limited US launch therefore remain part of the assessment, even if the idea already sounds interesting for many workflows.

Sources and documentation status: 25 October 2026.

Hamburg is our starting point.

The useful part travels with you.

Practical AI from Hamburg, explained for everyone.

Saved articles