Amin Ebrahimi closes Shahs of Kabob late, and the paperwork starts after that. He owns the restaurant with his family in South Florida, about 110 employees, and the invoices arrive store by store: produce, gas, payroll, the small charges that are easy to lose. Last year the work meant a browser with a hundred tabs open, and a phone that kept lighting up while he was on the floor.
This autumn he started handing the list to something else. He describes a task in a short message, and the work continues after he puts the phone down. Later a report is waiting. “It brings it all together,” he said. “I don’t need to open up 100 tabs.”
The product he is using is Claude Duet, which Anthropic opened as a research preview on October 21 and began pushing to a wider group of subscribers at the end of the month. It runs as its own app on iOS, Android and the web, and it is built on a different premise from the chat window that made the company’s name. The assistant messages you first.
Anthropic’s pitch is aimed at people who run something small. The company says it built Duet for professionals and small business owners with more on their plates than they can track, and priced it inside existing Pro, Max and Ultra subscriptions rather than as a separate product line.
From answers to errands
The mechanics explain the ambition. Duet keeps a virtual machine with its own browser, so a task does not need a laptop or a phone to stay awake; the model decides which applications to open and what steps to take, and notifies the user when a task is finished or needs review. It sends a short brief each morning across connected apps, and checks those accounts every few hours.
Checking continuously would burn through a plan’s usage limits, the company says, so the interval is a compromise between attentiveness and cost. Duet shares those limits with the rest of Claude, and background work counts against them.
That detail is where the business model shows through. Software has been sold by seat for as long as it has existed; an agent that works while nobody is watching is closer to labour, and metering it by the task is a different kind of invoice. A restaurant owner who delegates invoices, payroll preparation and a hiring post is buying hours, and the vendor’s cost is model turns.
Anthropic has been building toward that invoice for two years. Claude Code took the reasoning into software engineering, where the agent edits files and runs tests; Claude in Chrome put a browser under the model’s control; Cowork moved the work onto the desktop. Duet is the version that lives on a phone and starts the conversation.
The connectors do the rest of the selling. More than 6,000 plugins sit in the company’s marketplace: Gmail and Slack for missed messages, QuickBooks and Stripe for unpaid invoices, Gusto for payroll, Docusign for agreements, HubSpot for follow-ups, Shopify for the storefront. Payments run through Stripe’s Link wallet with a virtual card, so the real number is never handed over.
The approval problem
An agent that reads email and sends messages is making a judgment about where information belongs on every turn, which is a security problem before it is a convenience. Anthropic’s answer is a sequence of gates. Duet drafts and prepares on its own, but before it sends a message, makes a payment, signs a document or deletes anything, it shows the user what it is about to do and waits.
Some actions stay off limits even with approval, including trading securities and moving money between accounts. Credentials follow the design used for the company’s managed agents, so the model does not see passwords or sign-in tokens. A dedicated classifier and safety prompt check each message for the strictest tier of information, credit card numbers and government identification among them.
The gates are also an admission. Anthropic says the preview can misread a thread or misjudge what is urgent, that personal errands are weaker than work, and that Duet cannot yet show the user its browser. The company’s own researchers hand over a list before leaving for the evening and collect the results later.
That last pattern is the one that changes habits, because it moves work out of the hours when a person is at a desk. Boris Cherny, who leads the Claude Code team, dictates his to-do list on walks instead of saving it for later. Chelsea Larsson, who leads product behaviour design, hands over her list before leaving to be with her children.
Small owners were given early access, and their examples are more ordinary. A broker had a property showing flagged as scheduled on the wrong day, unprompted. A brand strategist had a client call prepared for her, and a clash with a flight pointed out before she noticed it.
The competitive question is whether proactivity is a feature or a category. Every large model company now ships an agent that can use a browser, and the differences between them are less about capability than about when the model decides to act and what it refuses to do alone. Duet’s answers are conservative: ask first, trade speed for a smaller blast radius.
For now the preview is limited to the United States and access runs through a waitlist, a sign the company does not yet know how it behaves for a hundred thousand owners with real invoices.
At Shahs of Kabob the invoices still get checked. Ebrahimi reads the summaries in the morning, corrects what the agent has misread, and lets it draft the rest. What he has bought is not intelligence, which he could already rent by the question. It is the hour after closing, when nobody is asking him anything and the tabs stay closed.
