Claude Duet Explained: Connected Apps, Approvals and Your First Workflow

What Claude Duet adds

Claude Duet is Anthropic’s proactive work assistant for connected apps, with background tasks and approval before consequential actions. Choosing which parts of your working day it can read is the first setup decision.

In its October 21 announcement, Anthropic describes Claude Duet as a proactive agent that briefs you across connected apps and continues tasks in a cloud virtual machine. It starts as a research preview through a waitlist for US Pro and Max subscribers, shares existing plan limits, and works on web, iOS and Android. Anthropic says it reads and prepares independently but waits for approval before sending messages, paying, signing or deleting. Apps can be disconnected.

Those are the vendor’s descriptions, not results from a CuratedAI hands-on test. Access and behavior may change during the preview. Check the announcement before planning a rollout.

Who should try it first?

A useful first candidate is someone with a repeated preparation task: collecting agenda items, finding unanswered questions or drafting a daily handover. The output can be checked against a small set of source records. A poor first candidate is a task where nobody can tell whether the result is complete until something has already gone wrong.

Start by writing down the current task in ordinary language. What inputs do you read? What do you produce? Who checks it? Which action actually changes someone else’s day? If those answers are unclear, connecting more apps will make the uncertainty larger.

Separate access from authority

Permission to read a mailbox and permission to send from it are different powers. A workflow may need the first without needing the second. This distinction is our setup recommendation, not a claim that every connector offers a separate read-only switch.

If a connector grants broader access than the task needs, an instruction to ignore unrelated records does not narrow that access. Choose a narrower account or integration where supported, use approved sample data, or postpone the connection.

A first workflow: prepare tomorrow’s meeting brief

Pick one meeting whose agenda and supporting material you already know. Use non-sensitive examples for the initial trial. Ask for preparation, then compare the draft with the original records yourself.

Prepare a brief for tomorrow’s product review.
Use only the meeting agenda and the project notes I identify.
List decisions needed, unresolved questions and source links.
Separate confirmed facts from your suggestions.
Flag missing or conflicting information instead of guessing.
Do not send messages, change the calendar or edit source records.
Stop after producing the draft for my review.

This prompt describes the desired workflow. It is not a security boundary and does not replace connector permissions or enforced approval controls. If the product cannot restrict the source set as requested, do not assume the wording made the restriction real.

Review three things: whether the source links support the claims, whether relevant items were omitted, and whether suggestions are clearly labelled. Record corrections while the source material is still fresh. Expand the task only after you can explain what improved and what still needs checking.

What a useful approval should show

Before approving a message, inspect the recipient, attachments and final text. For a calendar change, inspect the event, attendees, time zone and notification behavior. For a record update, inspect the old and new values. A summary such as “finish the follow-up” is too broad to judge.

When the proposed action differs from the task you delegated, pause and resolve the difference. Do not approve an unclear action just because the preceding draft was good. Preparing useful text and selecting the right destination are separate things to verify.

Why these controls matter beyond Duet

Walmart’s October 21 announcement describes Agent Profiles planned for the following month: separate agent purchase histories, customer-set approvals or spending limits, and revocation. Its Agent Gateway is being developed with a developer beta. These are announced capabilities, not a claim that everyone can use them today.

Glean’s headless agent builder announcement describes agents authored in files and run on its managed platform with enterprise permissions and an audit trail. That addresses a different question: how a team reviews and operates an agent over time.

Our interpretation is that useful delegation needs both a bounded task and visible authority. A personal assistant, shopping agent and enterprise runtime are different products, but the same evaluation questions travel across them: what can it see, what can it change, what must you approve, and how do you stop it?

Before you expand the trial

• Name the owner. One person reviews output and can stop the workflow.
• Inspect access. Check the actual connector grant, not just the task description.
• Check the result. Compare claims and omissions with source records.
• Inspect approval. Confirm the exact action and destination before proceeding.
• Set a stopping point. Define the deliverable and when to return for review.
• Check usage. Find out how background work affects your available plan limits.

For the broader context, read the October 21 AI briefing. For the next level of evaluation, see our guides to testing agents against prompt injection and controlling agent costs. Reviewed October 22, 2026, using the linked October 21 announcements.