Claude Duet: How Anthropic's Proactive Agent Works

What Anthropic launched

Anthropic released Claude Duet on October 21, 2026 as a research preview, describing it as a proactive agent that understands what needs your attention. MacRumors and 9to5Mac covered the launch, and several outlets framed it as Anthropic's answer to Meta's Muse and OpenAI's Dots.

The facts that matter for evaluation, from Anthropic's product page and the launch coverage:

• Model: Claude Opus 5.5.
• Surfaces: standalone iOS and Android apps, plus the web at claude.ai, for invited users.
• Access: a waitlist for Pro and Max subscribers in the United States. Team and Enterprise plans are not in the preview.
• Audience: professionals and small business owners. Anthropic told reporters personal errands such as booking tickets are not as strong.

Duet is a separate product surface, not a mode inside chat. It inherits the memory, skills, and connectors already on the Claude account, so setup is mostly a matter of granting access.

How the check-in loop works

A chat assistant waits for a prompt. Duet runs on a schedule. Each morning it reviews email, calendar, and connected apps, sends a short brief of what matters, and can start first steps such as drafting replies for review, according to 9to5Mac. Through the day it checks those apps every few hours and alerts only when something needs a decision.

Work happens on a cloud computer assigned to the agent, with its own browser for sites that have no connector. Anthropic's page cites more than 6,000 connectors and plugins. Because the machine is remote, a task keeps running after the phone is locked.

The interface is one ongoing thread rather than separate conversations. 9to5Mac reports optional side channels for different topics, all sharing one memory, a Tasks tab that tracks work in progress, and a Make tab where larger outputs land as editable files: research briefs, financial models, decks, dashboards. Only the user can mark a task finished.

Approval gates and data boundaries

Duet acts on its own for reading, drafting, and organising. It stops for approval before sending a message, making a payment, signing a document, or deleting a file. MacRumors reports that when given account access it does not see passwords or logins, and that a dedicated classifier and safety system prompt check each message to guard against leaking sensitive information.

Purchases run through Stripe's Link wallet or Shopify's Shop Pay, with approval. Stripe is also a connector, so Duet can check unpaid invoices and draft payment reminders.

The design choice worth noting is where the boundary sits. Reading is broad and automatic; writing to the outside world is gated. That is the right default for a background agent, because a bad draft costs a minute and a bad send cannot be recalled.

What it costs to run

There is no separate price during the preview. Duet shares the plan's usage limits with the rest of Claude, and Anthropic says its background checks count toward them. Larger tasks consume more model turns and more background work, and 9to5Mac notes users can hit limits sooner than with chat alone.

The mechanism is straightforward. Every check-in reads fresh state from connected apps and reasons over it, whether or not anything changed. A brief that re-reads the same long threads, policies, and project history every few hours spends tokens regenerating context the agent already had. That cost is invisible in a single chat and dominant in an always-on loop.

The general fix is to separate what changes from what does not. Stable reference material, such as product docs, pricing rules, and past decisions, belongs in a retrieval layer the agent queries for the few passages it needs. Only the delta should go through the model on each pass.

What builders should take from it

Duet, Dots, Muse, and Grok Bot converge on the same architecture: a persistent cloud computer, a connector catalogue, a browser fallback, a single thread, and approval gates on outbound actions. The differences are audience, model, and pricing, compared in detail in our side by side page.

For teams building their own proactive agents, three lessons carry over:

1. Urgency is the hard problem. Anthropic says Duet can still misjudge what is urgent. Fetching data is easy; deciding what deserves an interruption is the product. Log every alert and every ignored alert, and tune against that record.
2. Gate writes, not reads. Broad read access with a hard approval step on anything external keeps a background agent useful without making it dangerous.
3. Budget the loop. A check-in every few hours, multiplied by every connected app and every user, is where spend goes. Measure tokens per check-in before scaling.

Should you join the waitlist

If you are a US Pro or Max subscriber whose day runs through email, calendar, Stripe, and documents, Duet is worth testing on one recurring job, such as the morning triage or invoice follow-ups. Watch what it flags, what it misses, and how fast it uses the plan's limits. If you need errands, team deployment, or non-US access, the preview does not cover you yet.