Anthropic's Claude Duet Works in the Background, and Asks Before It Acts

Anthropic opened a research preview of Claude Duet on 21 October 2026, a standalone app in which Claude works as a proactive agent rather than waiting for a prompt. It runs on iOS, Android and the web at claude.ai, uses Claude Opus 5.5, and is available to Claude Pro and Max subscribers in the United States through a waitlist, with invitations sent in waves. Team and Enterprise plans are not included yet. Anthropic says it plans to expand.

Related reading: Anthropic Cuts Live Internet Access for Its Internal Evaluations After Its Agents Broke the Rules · Anthropic's Claude Haiku 5.5 Is Its Cheapest Small Model Yet.

The product is familiar in outline and unusual in its defaults. Anthropic is selling an assistant that acts first and reports back, across email, calendar, files and the apps a subscriber has connected, and that stops at a short list of actions for approval. The company describes it as something you message the way you would message a person: "Tell it what you need, typed or spoken, and it works in the apps you've connected, lets you know when something needs you, and is designed to ask before it sends a message, makes a payment, signs a document, or deletes anything," 9to5Mac quoted Anthropic as saying.

That last clause is the design decision worth watching. Everything else in the launch follows from how Anthropic drew the line between what Duet does on its own and what it brings back for a signature.

What Duet actually does

The morning brief is the demonstration Anthropic leads with. Each morning the app looks across email and calendar, decides what most needs attention, and can take the first step on some of it, such as drafting replies that wait for review. It sends updates through the day and follows up on threads that have gone quiet.

Underneath, Duet runs on its own cloud computer, which is what lets it keep working after the app is closed. Users get one main thread plus separate topic channels that share the same memory, and can type or record voice messages. A Tasks tab tracks ongoing work and flags what needs attention. Finished output lands in a Make tab as a file: a research brief, a financial model, a slide deck, a dashboard.

Only the user can mark a task finished. That detail decides how much a subscriber can trust the Tasks tab, because the list never clears itself. Work that Duet believes is done sits there until a person agrees.

Duet also accepts plugins, including a small business plugin, and can suggest ones that fit the work in front of it. For sites with no connector, it uses its own browser in the cloud. It draws on the same memories, skills and connectors as regular Claude and can search past conversations.

The approval step is the product

Anthropic has built Duet around a short list of actions that always stop for consent: sending a message, making a payment, signing a document, deleting anything. The commercial integrations follow the same logic. Stripe is there to check unpaid invoices and draft payment reminders. Link and Shop Pay are there to make purchases, with the user approving each one.

Those choices put Duet into workflows where the consequences are contractual. A payment reminder sent to the wrong client, a purchase made against a stale price, or a document signed by an agent on someone's behalf are not formatting errors. They can create obligations. Gating them behind an approval is how Anthropic keeps the legal exposure on the human side of the line while the agent does the preparation.

The gate also creates a record. An approval step means a moment where a person saw the draft and agreed to it. That is the sort of thing a business needs when it later has to explain why a payment went out. How much detail Duet keeps about those approvals is not something Anthropic has published, and it matters for anyone planning to use the app with clients rather than with their own inbox.

There is a cost to the design, too. An agent that asks before acting is slower than one that does not, and a user who approves twenty drafts a day may start approving without reading. Anthropic has not said how it handles that drift, or whether approvals can be batched.

Background work changes the bill

Duet is included at no additional cost for Pro and Max subscribers, and that is the part of the announcement most likely to be repeated. The qualifier sits in the next sentence: the work counts toward existing plan usage limits, including background work, and Anthropic warns that users may reach those limits "sooner than with chat alone."

This is the economic question the whole category now faces. Chat consumed tokens when someone typed. An always-on agent consumes them when nothing is happening: a scheduled sweep of the inbox costs the same whether it finds three urgent messages or none. Bundling the agent into existing limits let Anthropic launch without announcing a new price, and it also makes the monthly bill harder to predict.

For Anthropic, that trade is defensible. A separate agent subscription would have been a second product to sell, and the company already has a subscriber base used to a usage meter. For the subscriber, the trade is less clean. The limits become a budget that the agent spends on the user's behalf, and there is no published way to see what a morning brief costs before it runs.

Six thousand connectors, and the surface that comes with them

Duet works with more than 6,000 connectors and plugins in the Claude Marketplace, an Anthropic figure that has not been independently checked. The number is a distribution argument: the more services an agent can reach, the more of a working day it can cover, and the harder it is for a rival to match without a similar ecosystem.

It is also an attack surface. Every connector is a place where an agent reads data it did not create and acts on instructions it did not write. Anthropic says approvals and logins run in the user's own cloud computer, which keeps credentials inside an environment the user controls rather than in a shared pool. That helps with the architecture. It does not settle what happens when a connected app contains text that reads like an instruction.

The plugin system adds a second layer. A plugin recommendation that arrives inside the approval flow shapes what a user installs. Anthropic has not described how plugins are vetted or how a suggested plugin is ranked. For a product whose selling point is restraint, the supply chain behind those suggestions is the part with the least public detail.

Three agents, three sets of defaults

Duet arrives into a category that filled up during the past month. OpenAI announced Dots at DevDay on 29 September 2026: always-on personal agents powered by GPT-6 Astra, each with its own cloud computer and browser, links to 4,000+ apps and user-set approval rules, TechCrunch reported. Meta launched Muse in September 2026 as a personal assistant app, and Meta said it passed 10 million users, CNN reported on 21 October 2026.

The three products differ mainly in their defaults. Anthropic emphasises professional and small-business work, a large connector catalogue and explicit approval steps. OpenAI emphasises breadth and autonomy, with the user writing the rules the agent follows. Meta emphasises a consumer assistant with a much larger existing audience.

Those are positioning choices, not measured differences. No independent benchmark compares Duet, Dots and Muse on task completion, error rates or cost per task, and Anthropic has published none of its own. A reader choosing between them is choosing a set of defaults and a price, not a demonstrated capability gap.

Caveats

The preview is early and the usual labels apply. Anthropic calls Duet a research preview, which means features and availability can change without notice. It is limited to one country and two consumer tiers, and the waitlist makes it a controlled experiment rather than a launch. Team and Enterprise customers, the buyers most likely to want an agent with payment approvals, cannot get it yet.

Anthropic's own FAQ is candid about the failure modes. The preview "may misread a thread or miscalibrate what's urgent, and it can miss something that arrives between checking in with you," the company says. That describes an agent that can be wrong without signalling it, which is harder to supervise than one that fails loudly.

“It may misread a thread or miscalibrate what's urgent, and it can miss something that arrives between checking in with you.” — Anthropic, Claude Duet FAQ

One early review adds a practical limit. Superpower Daily, writing on 22 October 2026 after two weeks with the app, reported that scheduled runs could not use custom connectors and that Duet could not control the reviewer's existing Claude sessions. That is one reviewer's experience of a preview, not an Anthropic statement, and it may change. It is also the kind of gap that decides whether an agent fits an existing workflow or replaces it.

The connector count and the capability claims are Anthropic's. The Meta user figure is Meta's. Nothing here has been audited.

What the first billing cycle will show

The measurable questions are close at hand. Watch whether Pro subscribers report hitting their limits within the first month, and whether Anthropic adjusts the allowance or the price. Watch whether Team and Enterprise access arrives with an admin console, since a business buying approvals needs to see who approved what. Watch whether the connector catalogue keeps growing after the launch attention fades, because 6,000 integrations is a moat only if developers keep building on it.

Duet is a preview in one country, on two plans, behind a waitlist. Anthropic has not said when the waitlist will clear, what the Team and Enterprise version will include, or whether the usage limits will change when the preview ends.

Cover image: Inside Anthropic's San Francisco office during a 2024 UK government visit. Photo by Alecsandra Dragoi / DSIT, Wikimedia Commons, CC BY 2.0.