Technology — The chatbot era asked you to come to the AI. Anthropic's Claude Duet, opened as a research preview on Wednesday, reverses the direction: it watches your connected apps, sends a short brief each morning, keeps working on its own cloud computer while you are away, and messages you only when something needs a decision. It is the fourth major lab to ship an agent that reaches out first — and the same week, Walmart decided to let agents like it spend your money while Amazon slammed the door.
Anthropic announced Claude Duet on Wednesday, October 21, opening a waitlist for a research preview limited to Pro and Max subscribers in the United States. Duet is a standalone app on the web, iOS and Android, powered by Claude Opus 5.5. "It comes to you with what needs your attention, completes tasks you ask of it, keeps working when you step away, and checks with you before it acts on your behalf," the company said. You message it the way you would message a person — typed or spoken — in one ongoing thread, and it works inside the apps you have already connected to Claude, carrying over the same memory, skills and connectors, with no separate setup.
The timing is the story. Meta launched its Muse personal agent in September and told analysts on its earnings call this week that more than 10 million people in the US and Canada have set up Muse agents in about six weeks. OpenAI launched its always-on agents, called Dots, last month. Google has been folding agent features into Gemini. Anthropic is, by its own framing, late to this party — and it has chosen to enter it through the office door rather than the front door: Duet is built for professionals and small-business owners, people with inboxes, invoices and calendars that overflow, rather than for consumers booking dinner.
What Claude Duet actually does
The daily rhythm is deliberate. Each morning, Duet sends a short brief of the most important items across your connected apps. It then re-checks those accounts every few hours — Anthropic is candid that it does not monitor continuously, because nonstop checking would burn through a plan's usage limits much faster — and it messages you only when something appears to need attention. Background work counts against the same Pro or Max limits as ordinary Claude use, which the company acknowledges means heavy agent users will hit those limits "sooner than with chat alone."
The task range the company lists runs from the clerical to the substantial: scheduling meetings and drafting replies, checking QuickBooks or Stripe for unpaid invoices and drafting payment reminders, preparing payroll in Gusto for review, building dashboards with dozens of tools, producing financial or sales forecasts, and prototyping and building web apps. Anthropic says Duet can draw on more than 6,000 connectors and plugins in its marketplace, including Stripe, and can make purchases through Stripe's Link flow and Shop Pay — with approval. Tasks run on Duet's own virtual machine with its own browser, so work continues after you close the app or put your phone down. Inside Anthropic, the company says, Claude Code lead Boris Cherny uses Duet on walks to talk through to-dos, and the Duet team itself uses it to schedule user interviews and ship code.
The approval line: what Duet will not do alone
The guardrails are the product's defining choice. Before Duet sends a message, makes a payment, signs an agreement or deletes anything, it shows you exactly what it is about to do and waits for approval. Two categories are barred outright, even with approval: trading securities and moving money between your own financial accounts. A new classifier checks outgoing material for sensitive-information leaks; Anthropic says Duet never directly sees your passwords or sign-in tokens, and information relating to health or personal finances is shared only when you have explicitly told it to do so. The company also concedes the preview's limits plainly: Duet can misjudge what is urgent, it is tuned for work rather than personal errands, and it cannot yet show you its own browser.
The same week, retail split over agents that spend
Duet did not land in a vacuum. On Wednesday, Walmart unveiled the other half of the agent economy: an "Agent Gateway" that lets approved AI agents search its inventory and buy on a customer's behalf, with "Agent Profiles" arriving within the next month. Each agent you use gets its own profile inside your Walmart account — its purchases kept visibly separate from yours — and you choose the leash: approve every transaction, or let the agent buy on its own up to a spending cap such as $50, $100 or $200, deleting the profile at any time. Walmart compared the arrangement to adding an authorized user to a credit card.
Amazon is doing the opposite. The largest US online retailer blocks third-party shopping agents while promoting its own tools, including Alexa for Shopping and Buy for Me. Last month it blocked Meta's Muse from its site, telling users that unauthorized AI agents violate its conditions of use, and it previously sued Perplexity over purchases made through its Comet browser agent. The split is a clean map of incentives: Walmart, the challenger in e-commerce, gains if agents shop wherever they like; Amazon, the incumbent, loses its storefront's gatekeeping power if the customer who walks in is software it does not control.
Why this matters
Every major AI launch since ChatGPT has kept one assumption intact: the human starts. You open the app, you write the prompt, you wait. The agent launches of the past six weeks — Muse in September, Dots last month, Gemini's agent features, and now Duet — retire that assumption. The product is no longer an answer machine but a watcher: software with standing instructions, a memory of your work, and permission to act inside it. That is a different relationship, and a different risk. A chatbot can only be wrong when you ask it something. An agent that checks your accounts every few hours can be wrong on its own schedule — misreading a thread, misjudging urgency, preparing the wrong thing at scale — which is why Anthropic's approval line, and its refusal to let Duet trade or move money at all, may matter more than any feature list.
The numbers explain the rush. Meta's 10 million Muse setups in roughly six weeks is the fastest consumer-agent adoption figure any lab has disclosed, and Zuckerberg simultaneously called the product highly unprofitable — the classic shape of a land grab, in which the agent that lives in your phone first becomes the default intermediary for everything after. Anthropic's answer is to contest a narrower, richer niche: the professional whose employer already pays for software by the seat. Duet reaching Pro subscribers at the $20 tier — against Dots' current restriction to OpenAI's $100-and-up Pro plans — is a pricing wedge aimed at exactly that user. And the stakes behind the race are visible in Anthropic's own reported figures ahead of its anticipated IPO: revenues reported at $16.3 billion in the first half of 2026, up from $1.1 billion a year earlier, with net losses ballooning over the same period. Growth that steep buys many things; patience from investors about who owns the agent layer is not one of them.
Who wins, who is exposed
The winners, if proactive agents hold, are the platforms that own the connectors — the 6,000 integrations Anthropic cites are the real moat, because an agent is only as useful as the accounts it can touch — and retailers like Walmart that reprice their storefronts for machine customers early. The exposed are just as identifiable: standalone productivity apps whose entire value was watching one inbox or one calendar; Amazon's walled storefront, if shoppers come to prefer agents that compare across retailers; and users themselves, whose plan limits now fund a background process that spends them whether or not it finds anything worth reporting. Critics of the category make the blunter point that delegation has a cost no launch post lists: every task the agent prepares is a task you must review, and a morning brief you have to audit is a meeting by another name.
What happens next
Duet invitations will go out in waves from the waitlist, and Anthropic says Team and Enterprise plans — where the workplace money actually sits — are not included yet but are planned. The near-term tests are concrete: whether the preview's judgment about "what needs you" earns trust or trains users to ignore it; whether the usage-limit economics survive contact with real background workloads; and whether Walmart's Agent Profiles, due within a month, produce the first credible numbers on how much money Americans will actually let software spend for them. The agent race now has four serious entrants and one disclosed scoreboard — Meta's 10 million. Anthropic has bet that the winning agent is not the one that does the most, but the one that asks at the right moments. By the time Duet reaches general availability, that bet will have a measurable answer.
Related coverage: Anthropic cuts internet access for its AI evaluations · Stock market today: Nasdaq's worst day since June
Sources: Anthropic, Claude Duet announcement and @claudeai post, October 21, 2026; AFP, "Walmart welcomes AI agents as Anthropic launches its own," October 22, 2026 (via TechXplore); Engadget, "Claude Duet is Anthropic's answer to Meta Muse and OpenAI Dots," October 21, 2026; SiliconANGLE, "Anthropic introduces Claude Duet, a proactive AI agent for professionals," October 22, 2026; 9to5Mac and MacRumors, Claude Duet launch coverage, October 21, 2026.