Anthropic's Claude Duet Wants to Work Before You Ask
Most AI assistants wait. You type, they answer, and then everyone goes quiet until the next prompt. Anthropic’s newest product is built on the opposite instinct: it is meant to be already working when you look at your phone.
On October 21, the company behind the Claude models released Claude Duet, a standalone mobile app positioned as a personal agent rather than a chatbot. According to accounts of the launch, it keeps watch over the inboxes and calendars linked to your account, nudges you about bills that have gone unpaid, and can place orders and make payments after you give the go-ahead. It is available only to Pro and Max subscribers in the United States, and only after an invitation clears the queue.
The pitch is that the busywork of ordinary professional life — a meeting that needs rescheduling, a dashboard that needs rebuilding, a payment that needs chasing — should not require a fresh conversation every time. Duet runs in the background, spots the moment it can be useful, and acts, with human approval sitting on the transactions that matter.
That ambition runs straight into a wall that has little to do with model quality. Personal agents only know what the surrounding software lets them see. If a calendar, a bank or a shopping service declines to open its doors to an outside assistant, the agent goes blind. Several Chinese analyses of the launch made the same point: the apps are multiplying, but few platforms are eager to hand over the keys.
Anthropic is not alone in chasing this. OpenAI has been pushing its own agent-style products, and a wave of startups now sells assistants that book, buy and remind on a user’s behalf. The competition is shifting from whose model scores highest on a reasoning test to whose assistant actually finishes the errand — and who can persuade the rest of the software world to cooperate.
There is also the question of trust. An app that reads your mail, watches your schedule and can spend your money is asking for a remarkable amount of access. The approval step is the obvious safeguard, but the design assumes users will keep granting permissions that many have so far refused to give to any single company.
Duet arrives in a crowded moment. Nvidia used its Berlin developer conference this week to argue that AI is moving from answering questions toward running complete tasks, and Microsoft has been building oversight tools for fleets of agents inside companies. The agent era, in other words, is being assembled from several directions at once.
For now the rollout is deliberately narrow, invitation by invitation, in one country. The early weeks will show whether people want an assistant that taps them on the shoulder — and whether the apps it depends on are willing to unlock the door.