Claude Duet: Claude messages you first, and still asks before it acts

On October 21, 2026 Anthropic opened a research preview of Claude Duet, a separate app where Claude does not wait for you to type. It watches the apps you connect, tells you what needs your attention, and keeps working when you walk away. The announcement passed 11 million views, more than any other Claude post this autumn. This review covers what Duet actually does, the line it will not cross without you, who can get it, how people reacted, and what to watch for.

What Anthropic announced

The official post describes Duet in one sentence: Claude "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." As of November 4 it had about 11.4 million views and 55,000 likes.

The thread under it adds three points: Duet is built around Claude Opus 5.5, it carries over the memory, skills and connected apps from your Claude account, and it is an early research preview that "can still misjudge what's urgent".

How Duet works

Anthropic's announcement page describes four behaviours:

Behaviour	What it means in practice
It comes to you first	A morning brief of what matters across your connected apps, then a message during the day only when something needs you.
One ongoing thread	You type or speak in one conversation; a quick reply and a week-long project both start there.
Work continues when you leave	Duet has its own virtual machine with its own browser, so your laptop or phone can be off. You get a notification when a task is done or needs review.
Checks every few hours, not nonstop	Anthropic says checking constantly would use up your plan's limits too fast. You choose what you want to hear about and how often.

The tasks Anthropic lists range from scheduling meetings to building dashboards, sales forecasts, prototypes and small web apps. Duet can use the connectors and plugins in the Claude marketplace; the page names Gmail, Slack, QuickBooks, Stripe, Gusto, LinkedIn, Docusign, HubSpot, Canva and Shopify among them.

The examples on the page are deliberately ordinary: a property manager whose Duet noticed, unprompted, that a showing was booked on the wrong day; a brand strategist whose Duet prepared a brief for a client call and flagged that the call fell right before her flight.

The line it will not cross without you

An agent that acts on its own is only useful if you can trust where it stops. This is the paragraph that matters most on Anthropic's page:

• Asks first: sending a message, making a payment, signing a document, deleting anything.
• Never, even with approval: trading securities, moving money between accounts.
• Your passwords: Anthropic says Duet's credential handling follows the Claude Managed Agents design, so it "never directly sees your passwords or sign-in tokens".
• Sensitive data: a separate classifier checks each message, and Duet is designed to refuse to share the strictest tier of information, such as card numbers and government IDs.
• Disconnect anytime: you can remove any app Duet has access to.

Anthropic also lists what Duet does badly today: it can misread a thread or misjudge urgency, personal tasks such as booking tickets are less developed than work tasks, and you cannot yet watch its browser.

Who can use it, and what it costs in usage

Duet is a waitlist research preview. Six days after launch, Karan Ahuja, who leads the team that built it, posted the rollout order (about 200,000 views):

1. All US Max users come off the waitlist first.
2. Then US Pro users.
3. For people outside the US, the team will "share a date as soon as we have one".

Duet is included in the Pro, Max and Ultra plans and shares the same usage limits as the rest of Claude. Anthropic is explicit that the background checks count toward those limits and that big tasks, such as a financial model or a deck, take more model turns. If you already run close to your weekly limit, a proactive agent is one more thing drawing from it.

How people reacted

We looked only at the posts with the most reach (as shown on X on November 4, 2026). We left out the many posts that were only about rival products, in line with our rules.

Post	Reach	What it said
@claudeai (official)	11.4M views 55,518 likes	The announcement
@claudeai thread	1.29M views	Built on Opus 5.5; reuses memory, skills and connected apps
@claudeai thread	717K views	Research preview, waitlist, "can still misjudge what's urgent"
@ahuja_karan (Duet team)	200K views 1,350 likes	Rollout order: US Max, then US Pro, other countries later
@_catwu (Anthropic)	149K views 1,015 likes	"it'll notice work that needs to be done, and then do it before you ask"
@AlexFinn	127K views 1,237 likes	Hands-on after a few hours: "It's EXCELLENT"; praises the single thread that spins off tasks
@stripe	96K views 741 likes	Stripe says it is working with Anthropic so businesses can run inside Duet, including making purchases
@addyosmani (Anthropic)	65K views 463 likes	Duet runs on Opus 5.5 across 6,000 connectors and plugins

The most-viewed hands-on review came from Alex Finn, a few hours after launch:

The pattern: early testers praised the single thread that turns a stream of requests into a managed task list, and the fact that Duet proposed tasks on its own and asked before going ahead. The loudest complaint was not about the product but about access: US only, waitlist only. Some people pointed to self-hosted open-source agent projects while they wait.

Our honest review

What is genuinely good. The approval line is drawn in the right place. Reading, drafting and preparing are cheap to undo; sending, paying, signing and deleting are not, and those are exactly the actions Duet stops on. Running on its own machine solves the most annoying limit of agents in a browser tab: they die when you close the laptop. And reusing your existing memory, skills and connectors means the setup you already built in Claude carries straight over.

What to watch. A proactive agent spends usage even when you are not asking for anything, and Anthropic's own page says the background checks count toward your limits. Anthropic also says plainly that Duet can misjudge urgency, so the first weeks are about tuning what it may interrupt you for. Finally, every connector you add is another place where untrusted text can reach the agent; Anthropic says it uses its prompt-injection defences from browser use, but the safest connector is still the one you did not add.

What we did not test. We are outside the US and do not have access yet, so nothing here is from our own use. Everything comes from Anthropic's page and posts and the public reactions above, captured on November 4, 2026.

Verdict: the most ambitious Claude app so far, with sensible brakes. Worth joining the waitlist if your day is spread across email, chat and a few business tools; worth waiting for wider reports if you already live close to your usage limit.

How to prepare before your invite

• Clean up your Claude memory and skills now. Duet inherits them, so wrong or stale memories become wrong assumptions in its briefs.
• Connect only what you need. Start with one or two apps (for example email and calendar) and add more once you trust the briefs.
• Decide your interrupt rules. Write down what is worth a message (a client reply, a payment due) and what can wait for the morning brief.
• Read every approval card. The safety model only works if the "waits for your approval" step gets a real look.
• Watch Settings > Usage in your first week to see what the background checks cost you.

For the habits that make any Claude agent safer and more useful, see How to use Claude like a professional.

Sources: Anthropic's posts on X, the Introducing Claude Duet page and posts by Anthropic staff, captured November 4, 2026. Reach figures are as shown on X that day. Public reactions are summarised with links to the original posts; short quotes are the posters' own words. Availability and features can change after publication.