SDSignal Desk

Claude’s Duet agent is a little neurotic and surprisingly helpful

· Signal Desk newsroom

Analysis of reporting by The Verge. How we work

Image: The Verge

Anthropic's consumer agent shows its strongest use is catching what falls between email, Slack and docs, and its weakest is pretending to understand tone or do the creative work.

Why it matters

Claude Duet is Anthropic's AI agent for consumers, pitched at professionals and small business owners, and still a research preview limited to US Pro, Max and Ultra subscribers with a waitlist. The Verge's reviewer connected it to work email, calendar, Google Docs and Slack for a week.

The result is one of the clearer pictures yet of where agents help real people who don't write code. It can connect a meeting summary to a Slack message and spot a scheduling crunch nobody noticed. It also misreads nuance and offers to write things that should stay human.

From the desk

Our read is that Duet makes a good case for a specific kind of AI: a proactive backstop. The best moment in the review came when it noticed that upcoming time off would leave only a couple of working days before a big deadline, proposed a fix, offered to update the shared doc and drafted a message to the editor. That is useful work. It is the cross-checking that slips through the cracks when communication is spread across five apps.

The weaker moments are just as telling. Duet treated a soft promise to share a link as a hard deadline and nagged before the editor's workday had started. It picked up writing style well enough to mimic the reviewer's all-lowercase Slack habit, but it doesn't really grasp tone or what actually matters. Anthropic itself says it is still testing when the agent should speak up. The reviewer's description of a slightly neurotic assistant is fair, and we would add that an over-eager agent costs attention instead of saving it.

I'm also watching the privacy side, which the review raises honestly. Duet can see every email, draft and Slack message because the user opted in. Knowing that in principle is different from watching it work out your boss from a shared document. Multiply that across a workplace and you get questions about coworkers who never agreed to have an AI reading their messages, and about how much an employer or vendor can learn from that pile of context.

One detail we appreciate: messages sent through Duet carry a label saying they were sent with Claude. Disclosure like that should be the norm, not a nice touch. As agents draft more of our everyday communication, people deserve to know when they are talking to a person and when they are reading a machine's draft.

Where this leads if it scales is the real question. The reviewer wonders whether relying on Duet would leave her disconnected from her own work. That is the downside to take seriously. Agents that catch gaps make people better. Agents that take over the noticing, the remembering and the writing could make people worse at their jobs. Our advice to builders is to design for the first.

Context

Duet arrives in a crowded field of consumer agents the review mentions, including Meta's Muse and OpenAI's Dots. It sits beside Anthropic's Claude Cowork and Claude Code, and is designed as the most consumer-friendly of the three, with its own app and swipeable suggestion cards.

Who feels it

Knowledge workers
Proactive cross-checking across email, chat and docs can catch real conflicts, but users should expect misjudged urgency for now.
Small businesses
Owners without assistants may get the most from task triage, provided they are comfortable granting broad data access.
Coworkers and clients
People who interact with Duet users have their messages read by an agent without having opted in themselves.
AI product teams
The review is a roadmap: proactive gap detection is valued, while unsolicited creative drafting draws pushback.

What to watch

  1. Whether Anthropic widens access beyond the US waitlist and paid tiers
  2. Changes to how often and when Duet interrupts users
  3. Workplace policies on agents with access to shared Slack and email
  4. Whether the sent-with-Claude disclosure label becomes standard across rival agents

Read the original

Continue at the source.

The Verge

Companies: OpenAI, Anthropic, Meta

More on OpenAI

All OpenAI coverage →