Your assistant has started the meeting without you
For years, using a chatbot meant opening a box, typing something, and waiting for a response. The arrangement was clear: you started it.
Claude Duet changes the proposed relationship. Anthropic says the agent can examine connected work, identify what needs attention, and continue tasks while the user is away. It runs on a cloud computer rather than depending on your laptop remaining open. [Source 1]
The sales pitch is more free time. The operational question is who decides what deserves that time.
Software that answers a question helps with a decision. Software that chooses which question to bring you begins shaping the decision before you arrive.
That is useful territory. It also deserves better scrutiny than a cheerful animation of an inbox organizing itself.
The approval button is doing a lot of work
Anthropic says Duet prepares work independently but waits for approval before sending messages, paying, signing documents, or deleting material. Some activities, including securities trading and transfers between accounts, remain prohibited even with approval. [Source 1]
Those boundaries matter. They describe a product with limits, rather than a fictional digital employee empowered to run your entire life.
But an approval button cannot make a decision good by existing beside it.
Consider a hypothetical overdue invoice. The agent identifies the customer, finds correspondence, drafts a reminder, and asks permission to send. You approve. Only afterward do you discover that the customer disputed the invoice last week in a different conversation.
The final click was yours. The selection of evidence was the machine's.
Our interpretation: the quality of the approval request is as important as the existence of approval. A useful request must reveal what will happen, what supports it, and what might make it wrong.
Otherwise we have automated the paperwork and retained the liability. An impressively modern arrangement.
Approval records authorization. It does not establish that the decision is correct.
Disconnecting is not the same as forgetting
Duet inherits context from the existing Claude account. Its help documentation says disconnecting an app stops subsequent access, while previously acquired conversational memory remains. [Source 2]
People understand closing a door. They are less accustomed to closing a door on a system that remembers what was inside.
A user might reasonably want three different controls:
- Stop reading this account.
- Stop using information previously learned from it.
- Delete the retained information.
These are different requests. Product interfaces should make those differences legible without requiring the customer to develop a minor in information governance.
Memory can save enormous repetition. It can also preserve an outdated preference, an abandoned project, or a business relationship that has changed. The useful question is whether remembered context remains inspectable and correctable.
An assistant should remember that your accountant is Jane. It should also survive the news that Jane retired.
Proactive does not mean omniscient
Anthropic describes more than 6,000 available connectors and plugins on its product page. That is potential reach, not independent proof that every connected workflow performs reliably. [Source 3]
Its support documentation also makes important concessions: account checks occur at intervals, urgency can be misjudged, and users cannot currently watch or take over Duet's browser. Background activity consumes subscription usage limits. [Source 2]
These limitations change what sensible testing looks like.
A calendar conflict caught early can be valuable. A missed deadline can be expensive. Neither proves the whole product good or bad.
The proper unit of evaluation is a task under specified conditions.
Before handing over recurring work, compare:
- Time saved against time spent reviewing.
- Relevant alerts against interruptions.
- Completed tasks against tasks quietly abandoned.
- Correct summaries against omissions that change the decision.
A polished morning briefing is pleasant. A trustworthy record of what was checked, skipped, and uncertain is more useful.
The strongest case for letting it work
The strongest objection to all this skepticism is straightforward: humans already miss things. We forget invoices, lose attachments, and spend afternoons searching for documents we personally named.
Perfection would be a ridiculous admission requirement for software operating in a human office.
An assistant that reliably assembles drafts and flags ordinary problems could improve a small business even if it occasionally needs correction. Keeping consequential actions subject to review may be a sensible compromise between usefulness and exposure.
That case deserves to win when the evidence supports it.
What remains unproven by launch documentation is the balance: how often useful interventions occur, how much supervision they require, and whether that balance improves across sustained work.
We have not independently tested Duet. The company describes an early research preview, not a completed verdict on delegated judgment.
The chatbot leaving its window is a meaningful development. Before celebrating the empty chair it leaves behind, we should find out who is now deciding what belongs on the desk.
Frequently Asked Questions
Can everyone use Claude Duet?
No. The initial preview uses a waitlist for US Pro and Max subscribers. Team and Enterprise access is not included at launch. [Source 2]
Does approving an action prove it is correct?
No. Approval records authorization. The user still needs enough evidence to assess the proposed action.
Is this a hands-on review?
No. It is analysis of Anthropic's published launch and support documentation. Performance requires independent testing.
Sources and Methodology
Evidence reviewed October 22, 2026. Analysis of primary documentation; company performance claims are attributed. No independent product testing.
