Claude Duet checks with you before it acts. One reviewer says that makes it a nag.
Anthropic has run its proactive agent as a research preview since October 21. The Verge's first full hands-on test, published November 13, calls it "a decent helper, if a little jumpy at times."
Anthropic opened Claude Duet as a research preview on October 21. It’s a proactive agent: it sends a morning brief, keeps an eye on your connected apps and runs tasks on its own cloud computer rather than your laptop. Access is by waitlist, limited to US Pro, Max and Ultra subscribers.
On November 13, The Verge published the first long hands-on review we found. Allison Johnson tested Duet for about a week with her work email, calendar, Google Docs and Slack connected. Her verdict is mixed: Duet has been “truly useful” at times and “kind of a nag” at others.
What Anthropic says it won’t do alone
Anthropic’s rule is that Duet reads, drafts and prepares work on its own, but waits before anything leaves your hands. Before it sends a message, makes a payment, signs a document or deletes something, it shows you exactly what it’s about to do. A few actions stay blocked even with your approval, such as trading securities or moving money between accounts.
It checks your connected accounts every few hours. Anthropic says that interval exists to protect your plan’s usage limits, since checking nonstop would burn through them. A new classifier is meant to refuse requests to share the strictest tier of information, such as credit card numbers and government ID. Health and personal finance details go out only when you’ve explicitly told Duet to send them.
The launch page also admits the weak spot. Duet “might misread a thread or miscalibrate what’s urgent,” and Anthropic says it is still testing “its judgment about when to message you.”
What the one-week test showed
Johnson’s examples cut both ways.
| What Anthropic describes | What The Verge’s reviewer saw |
|---|---|
| Messages you only when something needs attention | One morning it pressed her about a deadline. Her draft had gone to editors late the night before, and Duet said before 9 a.m. that it still needed an edit. It missed that the deadline was not hard, and that it was only 7 a.m. where her editor was. |
| Spots what matters across your apps | It linked something said in a staff meeting, summarized in a Gemini email, to a Slack direct message. It also worked out who her boss was from a shared Google Doc. |
| Plans ahead and prepares work for review | She had missed that upcoming time off would leave two working days before a big draft. Duet proposed a fix, offered to update the Google Doc and drafted a Slack note to her editor. The note was tagged “sent with Claude.” |
| Helps with tasks you hand it | It kept offering to start her drafts. Johnson said that is “always a hard no” for her. |
The planning case is the one she calls “surprisingly helpful.” She also found the Slack draft “a little too good.” It copied her lowercase habit and shortened “signature” to “sig,” which she said she would never write.
Why the nag is the story
Johnson says that interacting with Duet “sometimes felt like interacting with a slightly neurotic personal assistant,” and the deadline example is the clearest case. Her concern was about tone. “A person can pick up on nuances in tone to understand what’s a big deal and what isn’t,” she wrote. She said Duet’s worry about the deadline was “a bit misplaced.”
My read: Anthropic’s approval gate is built for the risky actions, so nothing goes out unseen. It does little for the cost Johnson describes, which is attention. A morning brief plus check-ins is one more stream to sort through each day.
What this does not tell you
This is one reviewer, one week and one set of work accounts. The Verge doesn’t report how often Duet pinged her, how many suggestions she accepted, or how it handled errands. Anthropic says personal errands such as booking tickets are weaker than work tasks. Its three quoted early-access customers, Jess, Amin and Mya, are quotes the company chose to publish on its launch page. Team and Enterprise plans are not included yet, and the launch materials give no date for them.
For the same question about what an assistant does without asking, see our Google Docs add-on post, where asking before each edit is the default.
My read
My read: the approval gate is the right design for an agent that can read your email and send messages in your name. The harder problem is the one this review shows. An agent that interrupts you has to get the call about what’s urgent right most of the time, or each check-in becomes something to dismiss. Anthropic’s claim about judgment is the part to watch. Neither source gives a hit rate, and one week with one tester can’t produce one.
Note on sources
I didn’t use Duet myself. Everything about how it behaves comes from Anthropic’s launch article and product page, or from The Verge reviewer’s own testing. I read all three pages in full.