Claude Duet agent, in research preview for US Pro, Max and Ultra users, is useful but jumpy, The Verge finds

Claude Duet, Anthropic's AI agent for consumers, is a "research preview" available only to Pro, Max and Ultra subscribers in the US, with a waitlist to join, according to Anthropic. The Verge's Allison Johnson tested it for a week and found it useful at cross-referencing work information, but prone to mistimed nudges. Her review was published November 13, 2026.

What Duet is

Duet targets "professionals and small business owners" and is meant to use its judgment about what needs the user's attention most. Johnson describes it as more work-focused than Meta's Muse and OpenAI's Dots, and more consumer-friendly than Claude Cowork, which she says was more consumer-friendly than Claude Code. Like Muse, it has its own app. Johnson notes the app has been updated to fully support the iPhone Duo.

Once users connect productivity and chat services, Duet looks for ways to help and surfaces suggestions as cards. Users swipe left or right to dismiss them or have Duet start on them. Johnson likens the interface to a to-do list version of Tinder. Standalone pricing for Duet is not yet disclosed.

What the testing showed

Johnson connected her work email, calendar, Google Docs and Slack. Duet then:

• Linked a point from a biweekly staff meeting, summarized in an email from Gemini, to a direct message in Slack.
• Worked out who her boss is from a shared Google Doc used for weekly 1:1 meetings.
• Opened each morning with a brief message and highlights, including a reminder to pack an item for a video shoot before a flight.

In one case Duet caught a scheduling problem Johnson had missed: upcoming time off left only a couple of working days before another big draft was due. It proposed a solution, offered to update a Google Doc with weekly notes for her editor, and drafted a Slack message to him. The draft copied her habit of writing in lowercase and used an internal abbreviation for the CMS. It also shortened "signature" to "sig," which she says she never would. The message carried a "sent with Claude" label.

Where it misfired

Johnson had emailed an interviewee that a story would likely go up Tuesday. She sent her draft to editors late Monday night. Before 9AM Tuesday, Duet flagged that the story was still waiting for an edit. It missed that her message implied no hard deadline, and that it was 7AM local time for her editor.

Duet also often offered to start her drafts, which she declines on principle. At one point the task it offered to help with was reviewing itself. Johnson also says it is "kind of weird" knowing Duet can read every email, draft and Slack message she sends, though she opted in to that access.

Anthropic's blog post says the company is continuing to "test its judgment about when to message you," according to Johnson.

Verdict

Johnson says AI finding overlooked discrepancies across emails, DMs, meetings and shared docs is "really helpful." She prefers it as a backstop for her own organizational systems rather than a replacement, and does not want it in the creative process.

What this means

The review's strongest results came from Duet's access to several connected sources at once: tying a meeting summary to a Slack message, inferring a reporting relationship from a shared document, and spotting a calendar conflict. The weak point was timing and tone. Duet treated a soft commitment as urgent and ignored the editor's local time, which is the kind of judgment Anthropic says it is still testing. The "sent with Claude" label means recipients can see when a message was drafted by Duet. For prospective users, the practical question is how much of their email, calendar, documents and chat they are willing to expose. Duet's usefulness in this test scaled with that access, and the reviewer found the access itself uncanny.