Checked on 23 October 2026. Claude Duet, announced by Anthropic on 21 October, takes its assistant beyond a conversation that starts only when someone asks a question. The new application prepares work and continues in the cloud. However, the research preview starts through a waitlist for US Pro and Max subscribers, rather than a worldwide release.
The useful question is not whether this is another chatbot. It is how much work someone can delegate while retaining control over messages, information and decisions. The announcement matters because it makes the distinction between a model that answers and a product that maintains an assignment over time much more concrete.
Claude Duet: what Anthropic actually announced
The official Claude release notes date the research preview to 21 October. The product page confirms the initial US restriction, eligible subscriptions and invitation requirement. Joining a waitlist is not immediate access. The sources reviewed do not establish an Italian launch date or a timetable for worldwide availability.
That distinction belongs before any discussion of value for money. A subscription can be necessary without being sufficient. Equally, an impressive demonstration does not establish that the same setup is available to every account, in every country, with every connected application. Readers should check eligibility before changing an existing subscription.
| Question | Verified position |
|---|---|
| What is it? | A proactive agent in research preview |
| Who gets initial access? | US Pro and Max users through a waitlist |
| How is usage counted? | Against existing Claude plan limits |
| Sensitive decisions | Approval required according to Anthropic |
From a prompt to an ongoing assignment
Anthropic’s launch announcement describes an agent with a cloud environment that prepares material and contacts its user. This is not simply a cheaper model for API calls: that is a different question covered in our explanation of Claude Haiku 5.5 and its practical uses. Models, applications and permissions are separate layers.
Consider a small business asking for a summary of incoming customer requests. This is an illustrative workflow, not a test conducted by our newsroom. A sensible assignment defines the reporting period, permitted folders, priority criteria and expected output. Asking an agent to handle everything leaves too many consequential choices implicit.
Approval is the important boundary
Anthropic says actions such as sending messages, making payments, signing and deleting need confirmation. That is a stated product boundary, not evidence that mistakes become impossible. A useful evaluation should examine whether the approval screen clearly identifies the recipient, content, amount and consequences before the user accepts a proposed action.
For an initial deployment, separate preparation from external action. Producing a draft that can be checked carries a different risk from sending a customer a message. Even polished writing can contain an unauthorized commercial promise, an outdated delivery date or an attachment intended for somebody else. The confirmation should support substantive review, not a reflexive click.
What delegation really costs
During the preview, the service shares the subscription’s usage limits. Inclusion in a plan should not be read as unlimited continuous work. For eligible users, the relevant measure is the number of completed, reviewed assignments that fit normal usage, rather than the number of short answers an assistant can produce in a demonstration.
A reasonable trial records the assignment, time saved, corrections needed and notification frequency. If checking a summary requires rereading every original document, the benefit may shrink considerably. If an agent identifies a real exception and points to the right supporting material, its value becomes easier to establish without relying on a general impression of intelligence.
Connected information and the wrong context
Before connecting a mailbox or business system, reduce the scope of the first experiment. Non-sensitive sample documents help reveal what information is surfaced and what actions are suggested. This is our editorial recommendation, not a claim that every connector provides a particular isolation feature or that such testing guarantees a safe production deployment.
Legitimate access and appropriate use are different questions. An account may be allowed to read a file without that file belonging in an external reply. Instructions encountered in emails or web pages should not silently become new user orders. Checking the recipient and the information being shared matters precisely when the task looks routine and familiar.
A useful first trial
Start with a small, checkable result: a list of outstanding requests linked to their original messages, without sending replies. Set acceptance criteria beforehand: no missing customer, invented priority or incorrect date. Only after checking that result should the workflow gain another step, while retaining a clear way to stop work and revoke access that is no longer needed.
Claude Duet, dots and the wider agent shift
The direction resembles the change discussed in our OpenAI DevDay report on dots: assistants are becoming counterparts that follow assignments, rather than places to submit isolated questions. That resemblance does not prove which product performs better. A credible comparison needs matching tasks, access conditions and review requirements.
For readers outside the initial market, the practical conclusion is to watch geographic access, prepare bounded use cases and avoid purchasing a subscription on the assumption of an unannounced launch. Claude Duet is worth following for controlled delegation. Its real usefulness will depend on work completed correctly and efficiently, not simply the degree of autonomy described in a launch presentation.
