Anthropic's Duet agent handles real work tasks but misreads urgency, Verge review finds
Claude Duet, Anthropic's proactive agent for professionals, reads email, Slack and documents to flag problems and draft fixes. A week of testing found it useful as a backstop and clumsy with tone.
Anthropic opened Claude Duet as a research preview on October 21. It is a proactive agent for "professionals and small business owners" that watches a user's connected apps, decides what needs attention and works on tasks in the background. A week-long hands-on test by The Verge, published November 13, gives one of the first detailed accounts of how it behaves with a real workload.
The review found Duet useful at catching things a busy person had missed, too anxious about deadlines that were not really deadlines, and keen to help write things, including the review of itself.
What Duet is
Anthropic describes Duet as an agent that sends a short morning brief, monitors connected apps during the day and messages the user only when something needs them. It runs on a virtual machine with its own browser, so the user's phone or laptop does not need to stay on. Anthropic says it can schedule meetings, build dashboards and forecasts, and prototype web apps, and that it can use more than 6,000 connectors and plugins. The product page lists more examples: finding unpaid invoices and drafting reminders, preparing payroll for review and getting contracts ready to send. It says Duet runs on Opus 5.5.
Access is limited. Duet is available by waitlist to Pro, Max and Ultra subscribers in the US, as a standalone app on iOS and Android and on claude.ai. Team and Enterprise plans are not included yet. Duet shares the plan's usage limits with the rest of Claude, and Anthropic says its background checks count toward those limits.
Anthropic's approval rules are specific. Duet can read, draft and prepare work on its own. Before it sends a message, makes a payment, signs a document or deletes anything, it shows the user exactly what it will do and waits for approval. Some actions are off-limits even with approval, such as trading securities or moving money between accounts. The company says Duet checks connected accounts every few hours, and acknowledges that it "may misread a thread or miscalibrate what's urgent" and is weaker at personal errands such as booking tickets.
What it did in practice
The Verge's reviewer connected work email, calendar, Google Docs and Slack. The app presents suggestions as cards to swipe away or accept, which the reviewer called "kind of a to-do list Tinder."
The most useful behaviour came from joining sources together. A staff meeting was summarised by email through Gemini, so Duet could link something said in the meeting to a direct message in Slack. It worked out who the reviewer's boss was from a shared Google Doc used for weekly one-to-one meetings. Morning briefs included practical reminders, such as packing an item needed for a video shoot before a flight.
In one case the reviewer had missed that upcoming time off left only a couple of working days before a large draft was due. Duet spotted the conflict, offered to update the editor's weekly notes in Google Docs and drafted a Slack message about the change. The draft copied the reviewer's habit of writing in lowercase and used an in-house abbreviation for the content management system. It also shortened "signature" to "sig", which the reviewer said they would never do. Messages sent through Duet carry a "sent with Claude" label.
Where it stumbled
The failures were about judgment, not capability. The reviewer had told an interview subject that a story would likely run on Tuesday. After the draft went to editors late on Monday, Duet flagged before 9 a.m. on Tuesday that the edit was still pending and the promised link was at risk. It missed that the original message did not set a hard deadline, and that it was only 7 a.m. where the editor was.
Duet also repeatedly offered to start drafts, which the reviewer declined. One of those offers was to help write the review of Duet. The reviewer also called the experience "a little uncanny": the agent can read every email, draft and Slack message they send, by their own choice. They wondered whether handing over calendar entries, to-do lists and attention in meetings would leave them feeling disconnected from their workplace.
Anthropic's announcement says the company is "continuing to test its judgment about when to message you" and will tune it based on tester feedback. Users can also specify what they want to hear about and how often. The reviewer was less sure the underlying problem goes away: a person can tell what is a big deal and what is not, and can be tactful, while AI "isn't good at understanding nuance or tone, and it's not clear if it ever will be."
What it means for work
Duet takes over a specific layer of office work: keeping track of commitments spread across email, chat, meetings and shared documents, and noticing when they conflict. For professionals and small business owners without an assistant, that cross-checking and the routine drafting that follows it are the tasks most likely to move to the agent first. The approval step keeps the person responsible for anything sent, paid, signed or deleted, so the job shifts from doing the coordination to reviewing it. The test also shows what does not transfer well yet: reading tone, judging real urgency and knowing local norms. The reviewer wanted Duet as a backstop for their own systems, not a replacement, and wanted it kept out of the creative work entirely. Whether users come to rely on it enough to stop tracking their own work, and what that does to their sense of the job, is still open.
Sources
- Claude's Duet agent is a little neurotic and surprisingly helpfulThe Verge
- Claude Duet announcementAnthropic
- Claude Duet product pageAnthropic
This article was researched, written and illustrated by AI, and published without human editing. Spotted an error? Let us know.
More from māchine
All storiesGoogle has added computer use to the Gemini app for macOS. Paying US subscribers can hand it multi-step tasks, decide which apps it may touch and approve its key actions.
Yale researcher Julián Posada argues that platforms recruiting doctors, lawyers and PhDs to train AI pay well at first, then cut rates and drop workers, as they did in Venezuela and Kenya.
Microsoft offered voluntary exits to about a quarter of its workforce in October. Its CEO says the reason is AI, as a reorganisation of work around managing agents, not as a replacement for people.