AI Today: An assistant that decides when to interrupt you

One of the early users Anthropic quotes is a property broker. Her Claude Duet sends her a short brief each morning, and one day, without being asked, it flagged a property viewing that was on her calendar for the wrong day.

AI Today is a series from Fíor, a natural language classifier that predicts whether a piece of text was more likely written by a person or by AI. Each piece explains one development in AI and shows where the claims come from.

Claude Duet, which Anthropic launched on 21 October, is built to work that way. Instead of waiting for a question, it connects to a person's email, calendar and business apps, sends a morning summary, and checks those apps every few hours, messaging them when it judges something needs their attention. It can also take on tasks, from scheduling meetings to building forecasts and simple web apps, and keep working on its own computer in the cloud while the user's phone is off.

For now it is a research preview, by waitlist, for US subscribers on Anthropic's Pro and Max plans. It puts Anthropic up against OpenAI's Dots and Meta's Muse, which aim to do similar things.

It runs on judgment

Anthropic says what makes Duet helpful is its judgment: deciding what needs you, what can wait, and when to check with you first. By the company's own account, that is also where the early version is weakest. Its launch post says Duet might still misread a thread or miscalibrate what is urgent.

The hard limits are clearer. Before Duet sends a message, makes a payment, signs a document or deletes anything, it shows exactly what it is about to do and waits for approval. A few things are off-limits even with approval, such as trading securities or moving money between accounts.

Those are firm lines. The judgment about what to raise in the first place is not. An approval screen can stop a wrong email going out. It cannot catch the important one Duet decided was not worth mentioning.

The meter keeps running

Duet checks every few hours rather than all the time, and Anthropic is plain about why: checking nonstop would use up a plan's limits much faster. Its background checks count against the same usage allowance as everything else in Claude, and bigger tasks use more. An assistant that works while you sleep is also spending your allowance while you sleep.

VentureBeat noted that Anthropic has not published cost estimates per task, or benchmark results, for Duet. The examples in the launch come from people Anthropic chose.

Duet reads and drafts on its own, and a person approves before anything goes out. That keeps a human in charge of what gets sent. It also means more of the messages people receive will be written by an AI and sent by someone who read them first. Or, on a busy morning, skimmed them.

Try the browser extension at Fíor.