Claude Duet: Anthropic's Always-On Agent, Explained

Anthropic's launch post for Claude Duet, published October 21, 2026, describes a new app in which Claude "works as a proactive agent": it comes to you with what needs attention, completes tasks you assign, keeps working when you step away, and checks with you before acting on your behalf. The preview is included in Pro and Max plans at no extra charge, runs on Claude Opus 5.5, ships as standalone iOS and Android apps and on claude.ai, and starts with Pro and Max subscribers in the United States through a waitlist. Anthropic built it, in its words, for professionals and small business owners with more on their plates than they can track.

Duet joins a field of always-on agents that other labs opened this autumn, including OpenAI's ChatGPT Dots and Meta's Muse. This post explains Duet on its own terms: what it adds to the Claude app, where its approval rules stop, what it connects to, and what its background work costs against a plan's limits. The head-to-head for small businesses is in Claude Duet vs Gemini vs ChatGPT Dots for small business. Everything here comes from Anthropic's post, read on October 21, 2026; the product is a research preview we have not yet used.

The simplest description is the one Anthropic gives: Claude Duet comes to you first. The Claude app waits for a prompt; Duet starts the day with a short brief on the most important items across your connected apps, then messages you during the day only when something needs attention. Between those messages it checks your connected accounts every few hours. You talk to it by typing or by voice in one ongoing thread, so a quick reply and a week-long project start in the same place rather than in separate chats.

The second half is the background work. Duet has a virtual machine with its own browser, so a task does not depend on your laptop or phone being on. You describe the task in a short message, it works out which apps to use and what steps to take, and you receive a notification when it is finished or needs review. Anthropic's examples run from scheduling meetings and building dashboards to financial and sales forecasts, prototypes and web apps. The design centre, the post says, is judgment: deciding what needs you, what can wait, and when to check first. That is also why the model choice matters. Duet is built around Opus 5.5, Anthropic's current top general model, not a smaller routing model.

Claude already connects to apps, keeps memory, runs skills and can browse, and in Cowork it can already run scheduled tasks on Anthropic's servers with your devices off. What Duet adds is initiative: it decides for itself what needs you, without a task or a schedule being set first. The practical difference shows in Anthropic's own tester stories, which are Anthropic-selected and should be read as such. A real-estate broker had Duet flag a property showing that had landed on her calendar on the wrong day. A brand strategist had it notice that a client call sat right before her flight and offer to move it. Neither of those tasks was assigned; the agent found the conflict on a scheduled check.

The failure case is the mirror image. An agent that checks your inbox every few hours and messages you is only useful if its judgment about urgency is good, and Anthropic says in the same post that Duet "still might misread a thread or miscalibrate what's urgent", and that it is continuing to tune when it messages based on tester feedback. The post also says you can tell it what you want to hear about and how often. The setting to make first, then, is the scope of the checks, not the list of connected apps.

The approval rules are the most concrete part of the launch, and they are worth quoting as Anthropic states them. Before Duet sends a message, makes a payment, signs a document or deletes anything, it "shows you exactly what it's about to do and waits for your approval". A few things are off-limits even with approval; Anthropic's examples are trading securities and moving money between accounts. You can disconnect any app it has access to at any time. Credentials follow the design Anthropic uses for Claude Managed Agents, so Duet never directly sees passwords or sign-in tokens.

Around those rules sit the safety systems. Actions are approved by an auto mode classifier; defences against hidden instructions and prompt injection draw on Anthropic's browser-use research; and a new dedicated classifier with a safety system prompt screens outgoing messages for sensitive information. It is designed to reject sharing the strictest tier, including credit card numbers and government identifiers. Health and personal finance information is shared only on explicit instruction. Anthropic says the system was red-teamed, penetration-tested by an external firm, and is covered by its public bug bounty. The example to keep in mind: a payment goes through Stripe's Link wallet on a virtual card, so your real card number is never passed to the merchant, and Link asks you to approve the spend before it moves.

Duet uses the connectors and plugins already in the Claude Marketplace, which Anthropic counts at more than 6,000. The launch post's examples are the ones a small business would expect: find missed messages and draft replies in Gmail and Slack; check Intuit QuickBooks or Stripe for unpaid invoices and draft reminders; run payroll through Gusto; post a job on LinkedIn after your approval; send agreements for signature through Docusign; draft follow-up emails from a CRM; build social campaigns in Canva; and set up and manage a Shopify store, with a visual storefront editor described as coming soon. Purchases go through Link, Stripe's wallet, or Shop Pay at online stores. Signing in to sites with logins saved in 1Password is listed as coming soon, with your permission.

The example that shows the design is the invoice one. Checking QuickBooks for unpaid invoices is a read; drafting reminders is a write that stops at the approval step because a reminder is a message. The failure case to plan for is scope creep in the connections themselves: an agent that checks every few hours across every connected app will surface more than you asked for, and the fix Anthropic offers is to tell it what to watch. Connect the two or three systems where a missed item costs money, then widen. Our guide to Claude for small business covers the same connectors in the ordinary Claude app.

There is no separate price. For the research preview Duet is included in Pro and Max and "uses the same usage limits as the rest of Claude". The subtlety is in how it spends them. Anthropic says Duet uses those limits differently than a typical conversation: background checks count toward them, and larger tasks, such as a financial model or a deck, need more model turns and more background work. The every-few-hours cadence exists because checking nonstop "would use up your plan's limits much faster".

In plain terms, a Pro subscriber who also uses Claude heavily for writing or code is now sharing that allowance with an agent that runs while they sleep. The post gives no numbers for how much a check or a task consumes, so we print none. The example to run yourself: connect Duet, leave it for a day with the default cadence, and look at how much of your usage window it used before you assign it anything large. The failure case is the opposite order, handing it a forecast on day one and losing the afternoon's interactive limit to it.

The launch post is unusually direct about the edges, and they are the parts a buyer should weigh. Duet may misread a thread or miscalibrate what is urgent. Its core is built around work, so personal errands such as booking tickets "aren't quite as strong". It cannot yet show you its browser, which means that when a background task goes wrong you get the outcome and a notification, not a view of what it did on the way. And it lives in its own app for now, separate from the main Claude app, with the relationship between the two left open.

Each of those is a design limit rather than a bug, and each has a workaround for the preview. Keep delegated tasks to ones whose result you can check from the output, a dashboard or a draft rather than a completed purchase. Keep personal errands in the ordinary Claude app where you are in the loop. And expect the two apps to converge; Anthropic says early feedback will shape how they work together, which is the usual signal that the separation is a preview constraint, not a product decision.

Duet arrives into a category that filled up fast this autumn. OpenAI's ChatGPT Dots are always-on agents inside ChatGPT; Meta's Muse is a personal agent in its own app and in WhatsApp. We compared the field in our personal-agent comparison on September 30, before Duet existed, and will add a Duet row to it. The one-paragraph version of how Duet differs, from Anthropic's own description rather than from use: unlike Dots, it is a separate app rather than a feature of ChatGPT; and its approval boundary is stated as four actions that always wait plus refusals such as trading securities, where Dots let users write Custom Rules that allow, require approval for or block specific actions. All three are built on frontier models (Opus 5.5, GPT-6 Astra and Muse Spark) and run work on a hosted computer with a browser, so neither is a differentiator.

What it does not have yet is breadth of availability. Dots are rolling out to ChatGPT Pro and Business Premium in eligible markets, with enterprise tiers behind an administrator switch; Duet is US-only and waitlisted. For a small business choosing which to pay for, that is the deciding fact this week, and the detailed comparison on plans, apps and limits is its own post.

The waitlist costs nothing, so the real question is what to do with access when it arrives. The routing below assumes you are a US Pro or Max subscriber; if you are not, the honest answer is that the post gives you nothing to act on except the signal that Anthropic is building for operators, not only for developers, which it has been doing since Cowork reached web and mobile.

For businesses that want this class of agent wired into their own systems rather than through a consumer app, our AI transformation team builds the same approval-gated patterns on Claude's developer platform, where the limits and the audit trail are yours.