Claude Duet moves AI from answering prompts to watching the work

Anthropic has opened a research preview of Claude Duet, a standalone app designed to notice work, carry out tasks in the background and bring people in when judgment or approval is required. That sounds like a product update, but it marks a larger change in how AI enters an organisation: from a tool a person opens to an agent that keeps watching selected systems after the person steps away.

Anthropic’s announcement says the preview begins with a waitlist for US Pro and Max subscribers. Duet is available through iOS, Android and claude.ai, shares the user’s existing plan limits, and is separate from the main Claude app during the preview.

The important question is not whether Duet writes better chat replies. It is whether a proactive agent can monitor enough business context to surface the right issue, take useful preparatory steps and stop at the right boundary.

What makes Duet different

Duet is designed to initiate the interaction. It can prepare a morning brief from connected apps, monitor those apps during the day and notify the user when something appears to need attention. A user can also assign a longer task and let it continue on a virtual machine with its own browser while their laptop or phone is off.

Anthropic describes examples ranging from scheduling and inbox work to creating dashboards, preparing forecasts, prototyping and building web apps. The product can draw on Claude memory, skills and connected applications, and the company says its marketplace provides access to thousands of connectors and plugins.

That combination changes the interaction model:

• The agent holds a continuing thread of work rather than waiting for isolated prompts.
• It decides when information is important enough to interrupt the user.
• It can combine context from several systems instead of working inside one application.
• It can prepare work independently, then request a human decision at a defined boundary.

This is closer to delegated operations than a conventional assistant.

Approval remains part of the product

Anthropic says Duet asks for approval before it sends a message, makes a payment, signs a document or deletes something. Some actions, including securities trading and moving money between accounts, remain unavailable even with approval.

The company also describes controls for credentials, prompt injection and sensitive data. Duet uses an isolated virtual machine and is designed so the agent does not directly see passwords or sign-in tokens. A dedicated classifier and safety prompt are intended to prevent the strictest information—such as card numbers and government identification—from being disclosed, while other sensitive information requires explicit user instruction before it can be sent.

Those are meaningful design choices, but a research preview is evidence of direction rather than proof of reliability. Anthropic itself notes that Duet may misread a thread or misjudge urgency, and that background monitoring and larger tasks consume the same plan limits as ordinary Claude use.

What businesses should evaluate

The value of a proactive agent depends on the operating model around it. Before connecting a production account, a team should answer five questions.

1. What can it see?

Use the least privilege needed for a defined workflow. A broad connection to email, files, CRM, finance and messaging creates more context, but also a much larger disclosure and prompt-injection surface. Start with a narrow task and a small set of accounts.

2. What can it prepare, and what can it execute?

Separate read, draft and action permissions. A useful first deployment may let the agent read a shared inbox, prepare a proposed response and create a task, while requiring a person to approve every external message.

3. Who reviews its decisions?

An approval screen is useful only when the reviewer has enough context to spot a mistake. Define what evidence must accompany a request: original message, proposed action, affected system, destination, data being shared and any financial commitment.

4. How will the work be audited?

Record the instruction, sources consulted, steps taken, approvals received and final outcome. In a regulated workflow, “the agent did it” is not an acceptable explanation. The organisation still owns the decision and needs evidence that its controls operated.

5. What is the cost boundary?

Background checks consume usage, and complex tasks require many model turns. Decide how often the agent should monitor systems, what deserves an alert and when a task should stop. Otherwise an always-on helper can become an always-on source of noise and cost.

A sensible first pilot

Choose a workflow that is useful but recoverable: preparing a daily operations brief, finding unanswered customer messages, or assembling meeting context from approved sources. Keep external actions behind human approval. Measure missed items, false alarms, review time, completion quality and usage cost for several weeks.

The test is not whether Duet can complete an impressive demonstration. It is whether people can predict its boundaries, review its work efficiently and trust the evidence it provides. If the agent saves ten minutes but adds fifteen minutes of checking, the workflow has not improved.

Claude Duet points toward an AI product category that will matter well beyond one vendor: agents that observe, prioritise and act across the software a business already uses. The opportunity is real, but so is the governance burden. The organisations that benefit will treat permissions, approval design, audit trails and cost controls as part of the product—not as paperwork added after deployment.

If you are assessing where proactive agents fit safely, our AI consulting and integration work starts with the workflow, its risks and the human decision points.