Anthropic's Claude Duet shows the AI race shifting from building smarter models to deciding when agents act alone

Anthropic announced Claude Duet on Wednesday. The company describes it as a "proactive agent" that runs in the background, flags tasks to the user and handles a range of work. Business Insider reports that Duet is in research preview with a waitlist, so it won't immediately fight Meta's Muse or OpenAI's Dots for the mass market. It runs on Opus 5.5, Anthropic's flagship model, released in late September. Business Insider calls Opus 5.5 one of the most popular models in the industry.

The design follows a pattern that is quickly becoming standard. Each morning, Duet sends a brief on what is happening across the user's connected apps. It sends a message when something needs attention. It runs outside the user's computer, sends notifications and asks for approval on certain tasks. In a launch video, Mike Krieger, co-lead of Anthropic's Labs team, said the agent "can have the discernment and judgment around when to be proactive, when to act on your behalf." Product lead Karan Ahuja said on X that the team built it "for professionals and small business owners juggling a lot at work and at home." Ahuja ran product management at Meta Superintelligence Labs until July and joined Anthropic in August.

The launch was also aimed at competitors. Anthropic announced Duet on the day of Meta's earnings report and highlighted Ahuja's move from Meta. Meta noticed: Alexandr Wang, who heads Meta Superintelligence Labs, posted a meme about how similar Duet is to Muse. Meta also starts with a scale advantage. Apptopia estimates Muse had more than 14 million installs as of Wednesday. Apptopia also found that Claude users who installed Muse used Claude more than before. Even so, they spent fewer minutes a day in Claude than in Meta's app. And all of this comes as Anthropic heads toward an IPO that, according to Business Insider, could value it at up to $2 trillion.

Our reading: Duet confirms a thesis we have been tracking. The fight is moving from "who has the best model" to who controls the layer that sits between people and their digital lives: email, calendars, documents and notifications. We have argued that whoever owns this "plumbing" wins. A proactive agent is exactly that. It decides what you see first in the morning and what deserves your attention. Anthropic may have the stronger model, as our quality index suggests. Meta has distribution and a head start in installs. The Apptopia data cuts both ways. Agents don't seem to cannibalise each other (Claude usage went up among people who installed Muse), but attention time still tilts toward Meta.

The hard question isn't technical. It is how much the agent should decide for you. Krieger's "discernment and judgment" is a promise, not a demonstrated capability. Asking for approval "on certain tasks" is also weak protection if users approve almost everything out of fatigue. We have seen that with coding agents, where about 93% of permission requests get approved. An agent that reads all your connected apps and acts on your behalf needs least privilege by default (access to only what each task requires) and clear limits on what it can do without asking. That should be built in from the start, not added after the first incident. The research preview is the right time to prove it.

In the short term, expect friction. Agents will make mistakes, there will be privacy questions, and three companies will pull users in different directions. In the long term, the direction is promising. An assistant that handles the admin load of a professional or a small business, the paperwork and follow-ups that eat up the day, is exactly the kind of abundance we expect from AI: more time for the work that actually matters. Whether Duet delivers on that will depend on trust, and that will be measured in how often it gets things right, not in waitlists or memes.