Anthropic's Claude Duet Runs Your Work in the Background While You Sleep

Anthropic has introduced Claude Duet, a proactive assistant that monitors connected accounts, runs tasks in the background, and requests human input when an action needs review. The research preview extends Claude beyond user-initiated chats into persistent, hosted workflows that continue while a laptop or phone is offline.

Each morning, Duet compiles a brief from connected apps and flags items it considers urgent. During the day, it checks those services every few hours, drafts work, and sends updates through a single ongoing conversation. Users can assign longer jobs by text or voice.

Duet at a glance

A hosted agent that initiates work

Duet maintains one conversation and adds updates as tasks progress or new issues appear. It inherits memory, skills, and app connections from the user’s Claude account, which reduces setup and preserves context across sessions.

Background jobs run inside a virtual machine, an isolated cloud computer with its own browser. Duet determines which connected services to use, performs the necessary steps, and reports completion or requests review. The hosted environment allows jobs to continue when the user’s devices are off.

Duet polls connected accounts every few hours to limit model usage. Continuous polling would consume plan quotas faster, while the chosen interval leaves some delay between an external event and Duet’s response.

Opus 5.5 handles the triage

Anthropic built Duet around Claude Opus 5.5 because the product depends on judgment across several stages: identifying relevant events, estimating urgency, choosing tools, and deciding when to involve the user. Poor calibration could produce excessive alerts or allow important work to sit unnoticed.

The model also has to preserve context across applications. An email, calendar event, invoice, and travel booking may refer to the same commitment, so Duet must connect those records before proposing an action.

Approval gates and hard stops

Before Duet sends a message, makes a payment, signs a document, or deletes information, it displays the proposed action and waits for approval. Drafting, reading, and preparation can proceed in the background, subject to the product’s classifiers and permissions.

Some actions remain prohibited even after user approval, including securities trading and transfers between financial accounts. These hard limits prevent a user prompt or compromised workflow from authorizing the highest-risk operations.

Security for an unattended agent

Giving a hosted agent access to email, business software, and payment tools creates risks around stolen credentials, malicious webpage instructions, unauthorized actions, and sensitive-data leaks. Anthropic says Duet combines several existing defenses with controls developed for this product:

• Credential isolation follows the architecture used by Claude Managed Agents, keeping passwords and sign-in tokens outside Duet’s direct view.
• An auto mode classifier screens proposed actions for steps the user did not request. Anthropic uses the same type of control in claude.ai and Claude Code.
• Browser-use defenses detect prompt injection, including malicious instructions embedded in webpages or messages.
• Anthropic’s red teams and an external security firm tested the product, which is also covered by the company’s public bug bounty program.

A dedicated classifier and safety system prompt inspect messages for sensitive-information leaks. The strictest category, including credit card numbers and government identification, cannot be shared through Duet. Health information and personal financial details require an explicit instruction from the user before transmission.

Connectors supply the tools

Duet uses integrations from the Claude Marketplace to perform work across communication, finance, hiring, legal, and commerce services. Anthropic lists the following examples:

• Triage Gmail and Slack conversations, then draft replies.
• Find unpaid invoices in QuickBooks or Stripe and prepare reminders.
• Fill in hours, bonuses, and reimbursements in Gusto, then submit payroll after approval.
• Draft job postings and publish edited versions to LinkedIn.
• Prepare agreements and route approved documents through DocuSign.
• Set up and manage a Shopify storefront.

Business purchases can run through Link, Stripe’s digital wallet. Duet confirms the item, Link requests spending approval, and the transaction uses a virtual card that conceals the underlying card number.

Support for signing in with credentials stored in 1Password is planned. Anthropic says those logins will require user permission.

Early reports focus on caught mistakes

Anthropic’s early evidence comes from internal staff and a selected group of small-business owners, so the reports remain anecdotal. Their examples center on Duet detecting schedule conflicts and preparation work that a user had not assigned explicitly.

One broker reported that Duet noticed a property showing had been entered on the wrong day. A brand strategist said it prepared a brief for an upcoming client call, detected that the meeting was scheduled shortly before a flight, and offered to move it. Both examples illustrate the product’s defining behavior: raising a relevant issue without receiving a fresh prompt.

Preview limits meet quota pressure

Anthropic identifies several constraints in the current research preview:

• Duet may misread a conversation or misjudge urgency.
• Its workflows emphasize business tasks, leaving personal requests such as ticket booking less developed.
• Users cannot view the agent’s browser, which limits visibility into intermediate steps.
• Duet remains separate from the main Claude app.

Background work draws from the same usage allowance as ordinary Claude sessions. Periodic account checks consume quota, and complex assignments such as financial models or presentation decks require multiple model turns and tool calls. Pro users may therefore reach their limits faster than they would through chat alone. Anthropic has not announced pricing or separate usage policies for a wider release.

Patterns for agent builders

Duet gives Anthropic an inbox-like interface for an agent that initiates work and maintains a persistent relationship with the user. The company has applied related agent workflows in Claude Code and Claude in Chrome; Duet adapts those ideas for operators and small businesses.

Its architecture exposes four concrete patterns for developers building unattended agents:

• Persistent context: one conversation collects alerts, approvals, and completed work.
• Hosted execution: an isolated virtual machine keeps tasks running independently of user devices.
• Layered controls: classifiers, approval gates, credential isolation, and prohibited actions address different failure modes.
• Bounded polling: scheduled checks balance responsiveness against model usage and cost.