At seven in the morning it sends a short brief. Through the day it watches a calendar, an inbox and whatever else has been connected, and it messages only when something needs a decision. Anthropic calls the product Claude Duet, and it opened on Wednesday as a research preview, by waitlist, for Pro and Max subscribers in the United States.
Duet is a separate app — iPhone, Android and the web — rather than a tab inside Claude. It inherits the memory, saved skills and connected accounts of the main product, and the whole relationship runs through one thread that accepts typing or voice, so a one-line reply and a weeklong project start in the same place. Anthropic built it around Claude Opus 5.5 and included it in existing Pro and Max plans, with the same usage limits as the rest of Claude. There is no free tier in the preview.
The tasks it advertises are the ordinary work of running something small: scheduling meetings, assembling dashboards that pull from dozens of tools, drafting financial or sales forecasts, prototyping, building a simple web app. The brief is the hook. Anthropic’s bet is that most people do not know what to ask an assistant, and would rather be told what needs them.
The company’s own testers describe a division of labor rather than a replacement. Boris Cherny, who leads the Claude Code team, talks through his mental to-do list with Duet on walks. Chelsea Larsson, who leads model and product behavior design, hands over her list before leaving to be with her children and reads the results in the evening.
Jess Provencher, a broker and owner of Pro Homes, said Duet flagged a property showing that sat on her calendar for the wrong day without being asked. Amin Ebrahimi, who owns Shahs of Kabob in South Florida with about 110 employees, uses it for accounting, customer support and marketing, store by store.
Those examples point at the target customer: a professional or an owner with too many open loops and no operations staff. Small businesses buy software differently from large ones. There is no procurement committee, no security review and no integration team, which makes the segment easier to enter and harder to keep, because the same buyer will switch if the product stops being useful for a week.
The harder engineering is not the tasks. It is the judgment about when to interrupt. An agent that pings too often gets muted; one that stays quiet misses the thing it was hired to catch. Anthropic says Duet decides what needs the user, what can wait and when to check first, and that actions pass through an approval classifier before they run.
Anthropic has also published an unusual amount of detail about what the agent cannot do. Duet never sees a user’s passwords or sign-in tokens, following the design used for its managed agents, and it inherits the company’s defenses against hidden instructions in web pages and email. Credit card numbers and government identification are blocked from being shared by a dedicated classifier and system prompt; health and personal finance information pass only on explicit instruction. Anthropic says red teams and an outside firm ran penetration tests, and the app is covered by its public bug bounty.
That disclosure is partly a sales document. An assistant that reads email and sends messages on a user’s behalf is the most attractive target in the product line, and the first serious prompt-injection incident at a consumer agent will be blamed on whoever shipped it first. Getting the permission model right is now a marketing asset.
The competition arrived quickly. Meta’s Muse, launched in September, has been set up by more than 10 million people in the United States and Canada, Mark Zuckerberg said on Wednesday’s earnings call. OpenAI’s dots, introduced Sept. 29 and built on GPT-6 Astra, work across Slack and Teams and are being offered to Pro subscribers outside parts of Europe. Both are always-on agents that pursue goals without a prompt.
The difference is distribution. Meta can put an agent in front of billions of people who already open its apps every day. OpenAI can do something similar through ChatGPT, which the company says 1.2 billion people use weekly.
Anthropic sells subscriptions and models, and its consumer surface is smaller. Duet’s job is to make an existing subscription worth more, which is a margin argument rather than a growth argument.
The cost side is less discussed. Monitoring connected apps burns tokens without a request attached, and a flat subscription with shared usage limits is a wager that the value of being told what matters exceeds the inference bill. If heavy users treat the brief as a subscription to their own life, the limits will bind.
Anthropic has been meeting prospective investors ahead of a widely expected listing, and a consumer agent with a waitlist is a useful story about recurring revenue rather than model benchmarks. It is also a story that will be tested by retention, not signups.
The open question is behavioral. Search trained people to ask. Chat trained them to ask more cheaply.
Duet assumes they would rather be told, and that a software product can earn the right to interrupt by being right often enough. The mute button is the real reviewer, and it will render a verdict long before the waitlist clears.
