Claude Duet: The AI Agent That Works Before You Ask

By Muzamil Ahmad · Updated 2026-10-22 · Chatbots

#chatbots#claude duet#anthropic claude duet#claude duet waitlist#proactive ai agent#claude opus 5.5 agent#duet vs muse vs dots

Save

Every AI assistant you have used until now has one thing in common: it waits. You open the app, you type, it answers, you close the app, it stops existing. On Wednesday, October 21, Anthropic shipped the opposite of that. Claude Duet is a standalone agent app that comes to you first: it reads your email and calendar overnight, sends you a morning brief on what actually needs you, keeps checking your connected apps every few hours, and does real work on its own cloud computer while your phone is in your pocket and your laptop is shut.

Anthropic's own summary of the idea is the clearest one: Duet "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." It runs on Claude Opus 5.5, Anthropic's flagship model. It is a research preview, it is waitlist-only, and for now it is limited to Claude Pro and Max subscribers in the United States. It is also, unmistakably, Anthropic's entry into the agent war that Meta started with Muse in September and OpenAI joined with Dots at DevDay — the race to build the always-on personal agent. This guide covers what Duet does, how it works, who can get it, how it compares with Muse, Dots and Gemini, and the limits Anthropic itself is warning about.

What Claude Duet actually is

Claude Duet is a separate app — iOS, Android and the web at claude.ai — not a mode inside the normal Claude chat window. The interaction model is messaging, not prompting: you talk to it the way you would message a very capable assistant, by text or voice, in one continuous conversation. Three things make it different from the chatbot you already know.

First, it initiates. Every morning it looks across your connected email and calendar, picks out what it judges most needs your attention, and sends you a brief — with replies already drafted for the messages it thinks you should answer. Between briefs, it re-checks your connected accounts every few hours (you choose what it monitors and how often) and flags things like scheduling conflicts or a thread going stale.

Second, it persists. Duet runs on its own virtual computer in the cloud, with its own browser. Close the app, turn off your phone, go to sleep — an assignment you gave it keeps running, and it messages you when the work is done or when it hits something that needs a decision.

Third, it uses your existing Claude setup. Your memory, skills and connected apps carry over from your Claude account, so there is almost nothing to configure. Through the Claude Marketplace it can act inside more than 6,000 apps and plugins — Gmail, Slack, QuickBooks, Stripe, Gusto, LinkedIn, DocuSign, HubSpot and Shopify are the names Anthropic leads with.

How Duet works, step by step

The morning brief. This is the product's centerpiece. Instead of you opening five apps to reconstruct your day, Duet opens them for you: unpaid invoices sitting in QuickBooks or Stripe, a contract waiting on your signature, a calendar collision, the email thread that went quiet two days ago. The brief arrives with drafts attached — the reminder email is written, the reply is composed — so your job shrinks from doing the work to approving it.

Background monitoring. Duet checks connected accounts on a cycle of a few hours and surfaces what changed. You control the scope: which apps it watches, and how frequently. This is also where the first honest cost lives, and we will come back to it: those background checks consume your plan's usage allowance whether or not they find anything.

Task execution on its own computer. Assignments arrive as messages — "build me one page I can check every morning" is one of Anthropic's own examples — and can be anything from scheduling a meeting to building a dashboard, a financial forecast or a small web app. Because Duet has a cloud machine with a browser, it can do multi-step work across tools rather than only answering inside a chat.

The approval gate. Anything consequential stops and asks. Duet will not send a message, make a payment, sign a document or delete a file without your explicit go-ahead. Payments run through Stripe Link and Shop Pay when you approve them. Anthropic adds two safety details worth knowing: the approvals, logins and safety checks happen inside your own cloud computer, and the company says Duet does not directly see your passwords. A dedicated safety classifier and system prompt are meant to block the most sensitive data exposures.

The work Anthropic is showing off

Launch examples are always chosen to flatter, so read these as the shape of the product rather than proof — but they are specific, and they are aimed at small business owners and professionals rather than demo audiences. On Anthropic's Duet page, the agent is shown catching a county health permit fee that needs a $640 approval, redlining a contract renewal to Net 60 payment terms with the changes already summarised for review, and finding 27 wrong delivery orders with the refunds already drafted. The recurring pattern in the examples is bookkeeping-grade work: chasing unpaid invoices, preparing payroll in Gusto for a human to review, building the recurring brief or dashboard a owner checks each morning.

If that list sounds familiar, it should. It is the same "AI that does the boring follow-up" promise we scored highly in our best AI tools for small business guide — except here the follow-up happens before you remember the invoice exists.

Who can get Claude Duet, and what it costs

The gate is narrow, deliberately:

• Plans: Claude Pro and Claude Max subscribers only. Team and Enterprise plans are not included at launch, and Anthropic has not announced when they will be.
• Region: United States only.
• Access: a waitlist for the research preview. Invitations go out in waves; joining the list does not mean access today.
• Price: there is no separate Duet price. It is bundled into the existing Pro and Max subscriptions at no additional cost — which is a genuine contrast with Meta, which has tested premium agent tiers at far higher prices (we covered that pricing in our Muse Wonder guide).
• The catch: Duet shares your plan's existing usage limits. Background checks count toward them, and complex assignments consume more. An agent that works all day on a meter designed for chat is the equation every early user will have to watch.

Claude Duet vs Muse vs Dots vs Gemini

Duet is the fourth serious entry into a category that formed in about six weeks. Here is the landscape as of October 22, 2026:

Muse won the distribution race before it started: it rides Instagram, WhatsApp and Facebook, and it reportedly climbed the download charts within days. Its question is depth, not reach — downloads are not completed tasks.

Dots is OpenAI's persistent agent, launched at DevDay on September 29 with its own cloud computer and browser and a reported 4,000-plus connected apps. OpenAI is also spreading its models into everyone else's platforms — its models just landed on Google Cloud, with agent tooling like Google's Antigravity to follow, as we covered in OpenAI models on Google Cloud.

Gemini is the quiet incumbent. Google has not shipped one branded agent app; it has been threading agent features through products billions of people already open every day, and its models keep absorbing agent benchmarks — four Gemini 4 Argon agents finished Minecraft unaided in the run we covered in our Gemini 4 Argon story.

Duet's bet is different from all three: the smallest launch, the strongest model underneath it, and the strictest approval gate. Anthropic is spending its flagship Opus 5.5 — the model it positions against GPT-6 Astra on agentic benchmarks (the same frontier tier we compared in ChatGPT vs Claude vs Gemini) — on judgment calls: what is urgent, what can wait, and when to stop and ask. Whether that caution reads as trustworthiness or as being late is the argument the next few months will settle.

The honest limits

Anthropic deserves credit for printing its own caveats at launch, and they are not small print:

• It can misjudge urgency. Anthropic says plainly that Duet "can still misjudge what's urgent" and may misread a thread. An agent whose core job is deciding what deserves your attention will sometimes decide wrong — flagging the trivial, missing the important. Expect to tune it, and expect some odd pings early on.
• It burns your existing limits. Every background check draws on the same Pro or Max allowance you use for chat and Claude Code. Heavy users may find the always-on agent competing with their own usage — Anthropic has not published how much a typical day of Duet costs against the meter.
• Businesses mostly cannot have it yet. The customers with the deepest agent workloads — Team and Enterprise accounts — are excluded from the preview with no date given. For a product whose best examples are invoices, payroll and contracts, that is a real gap, not a footnote.
• One country, one language of rollout. US only at launch. Everyone else waits, including markets where Claude otherwise competes hard.
• Personal errands are the weak spot. Coverage of the launch notes Anthropic steering expectations toward work tasks; booking-style personal errands are reportedly not its strength yet.

Independent coverage has been broadly positive on the design and cautious on the scope — TechStrong's launch report stresses the approval model and the SMB examples, while Silicon Noon's analysis makes the sharpest point about the whole category: an agent that acts for you needs standing access to your accounts, so the real product being sold is trust in when it asks and when it does not.

What this means

Step back from the launch and the pattern is the story. In roughly six weeks, Meta, OpenAI, Google and Anthropic all arrived at the same product: a persistent agent with its own computer, your apps connected, working between your prompts. The chatbot era's question — "which model answers best?" — is being replaced by a harder one: "which agent do you trust to act while you are not looking?"

That shift rearranges the competitive axes. Raw model quality still matters (it is why Anthropic put Opus 5.5 on triage duty), but distribution, integrations and the approval design now matter as much. Meta has the audience. OpenAI has the app count and, as of this week, a seat inside Google's cloud. Google has the surfaces. Anthropic has the most conservative gate and the most candid fine print — a coherent position for the lab that has made safety its brand, and a slower one.

For Claude subscribers in the US, the practical move is simple: join the waitlist, and when access lands, start Duet on read-and-draft work — briefs, monitoring, drafted replies — before you let it anywhere near payments or signatures. For everyone else, Duet is a preview of what your existing assistant, whichever one it is, will look like by next year.

Verdict

Claude Duet is the most carefully designed of the always-on agents and the least available — flagship model, strict approvals, honest caveats, tiny door. If you are a US Pro or Max subscriber, it is the easiest yes on this page: it costs nothing extra to try. If you are anywhere else, or on a business plan, the agent war still has a seat saved for you; it just is not this one yet.

The bottom line: Duet will not be the agent most people use first. It might be the one that defines how much an agent should ask before it acts — and in a category built on trust, that is not a bad thing to define.