Claude Duet Explained: An AI Agent That Works While You’re Away

01 / The story

What is Claude Duet?

Claude Duet is an AI agent that can bring work to your attention instead of waiting for every instruction. Anthropic opened its research preview on October 21, 2026. The company describes a system that prepares a morning brief, checks connected work information, completes assigned tasks in the background and contacts you when it needs a decision. [1] [2] [4]

The central change is the starting point. A normal chat begins when you ask a question. Duet can notice a message or task in connected context and suggest what needs attention. That does not mean it knows everything happening in your business, or that it should make important decisions without you. Its access and judgment have limits. [2] [3]

Anthropic says it designed Duet around Claude Opus 5.5. The Help Center says users cannot choose a different model in Duet. A model is the underlying AI system; Duet adds connected apps, background work and an approval process around it. Our AI-agent explainer separates the model from the workflow it operates. [2] [3]

02 / The story

Watch: Introducing Claude Duet

Video in this section · scroll down to watch

The following video is posted by Claude’s YouTube channel. It is product communication from the publisher, not an independent Qlauson test. Use the written source record below to check specific features and preview conditions. [5]

03 / The story

How does Claude Duet work while your computer is off?

Duet carries over memory, skills and connected apps from your Claude account. Skills provide reusable instructions for work, while connectors give the system access to supported services. Anthropic says Duet runs in a virtual machine with its own browser, so your laptop or phone does not need to stay on. This is cloud execution, not an agent secretly operating your local computer. [2]

Its checks are periodic. The announcement says connected accounts are checked every few hours to conserve usage; it does not promise continuous monitoring or a fixed minute-by-minute schedule. A message arriving just after a check can therefore be missed until later. Do not use it as an emergency alert system. [2]

A practical sequence is: connect an approved app, explain the outcome you want, let Duet prepare or perform permitted work, inspect any approval request, and check the finished result. The final step matters: a completed task can still contain an incorrect interpretation, missing attachment or wrong assumption. These are review recommendations, not claims that Qlauson tested the sequence.

04 / The story

What can Claude Duet do for everyday work?

Anthropic’s examples include morning briefs, meeting scheduling, dashboards, forecasts, presentations, prototypes and web apps. The product page also describes finding unpaid invoices and drafting reminders, preparing payroll for review and getting contracts ready to send. Treat these as documented use cases, not a guarantee that every account or document will work correctly. [1] [2]

For example, an invoice reminder is useful only if the agent has identified the right invoice, payment status and customer. A payroll preparation task should end with a qualified person checking the figures and authorization. A draft contract is not legal approval. Preparing an action and taking responsibility for it are different jobs.

Start with low-impact preparation: a brief, a draft response or a list of records to inspect. Move to more consequential tasks only when the organization understands the permissions and approval process. Our agent approval guide discusses that boundary in more detail.

05 / The story

What does the 6,000-plus connector claim actually mean?

Anthropic says Duet can use more than 6,000 connectors and plugins in the Claude Marketplace. The announcement names workflows involving Gmail, Slack, QuickBooks, Stripe, Gusto, Docusign, HubSpot, Canva and Shopify. That is a company-reported count, not an independent Qlauson audit. [2]

A connector is a link to a service; it is not a promise that every function in that service is available. Existing account permissions, connector restrictions and the task itself affect what can be done. The Help Center specifically notes limitations with some connectors and says Microsoft 365 is not available yet. [3]

Do not confuse announced future integrations with current ones. The launch article labels saved-logins access through 1Password and a richer Shopify visual storefront editor as coming soon. A planned feature is not a reason to grant access today. Our MCP explainer offers background on connecting AI tools without treating every connection as equivalent. [2]

06 / The story

Which actions need human approval?

Anthropic says Duet is designed to show a proposed action and wait for approval before sending a message, making a payment, signing a document or deleting information. This is an important safeguard, but “designed to” is not an independently proven guarantee of perfect enforcement. Read the proposal, not just the approval button. [2] [3]

Some actions are prohibited even if the user approves them. The Help Center lists trading securities, moving money between accounts and accessing someone else’s account. Approved purchases through supported services are a different category; a blanket statement that Duet can never make any purchase would be inaccurate. [3]

A useful approval check has four parts: the destination, the information being sent, the financial or legal consequence and the ability to undo it. If any part is unclear, stop and ask for a revised proposal. Automation should shorten preparation, not make accountability disappear.

07 / The story

Is Claude Duet safe to connect to work accounts?

Anthropic says Duet never directly sees passwords or sign-in tokens. Its Help Center describes a separate sign-in form that sends credentials to the browser rather than putting them in the chat. The announcement also describes a classifier and safety instructions intended to restrict disclosure of highly sensitive information. These are attributed architecture and safety claims, not proof that no sensitive data is processed. [2] [3]

Prompt injection remains relevant. This is when hostile instructions in a page, message or document try to redirect an AI system away from its legitimate task. Anthropic describes testing and defenses, but its Help Center explicitly acknowledges that no defense is perfect. Incorrect interpretations and excessive permissions create risks even without an attacker. [2] [3]

Disconnecting an app is not the same as erasing all past context: the guide says memory from earlier conversations is kept. Review account memory and applicable privacy settings separately. Before connecting company information, check internal policy, retention requirements and who is allowed to authorize actions. Do not assume consumer preview eligibility establishes organizational compliance. [3]

08 / The story

What are the preview’s practical limitations?

Anthropic acknowledges that Duet can misjudge urgency, misunderstand message threads and miss events between checks. Personal errands are less developed than work tasks. The current guide says you cannot watch or take over its browser, even though you can see task-related activity and communicate with the agent. Browser visibility and task visibility are different things. [2] [3]

The guide also says Duet cannot see all projects or local computer files, does not make phone calls and does not send text messages to other people. Approved email through connected accounts is supported. These restrictions matter when choosing a workflow: an agent cannot complete a task simply because a conventional employee could. [3]

Keep a fallback for important work. Check the actual calendar booking, delivered message or updated record after a task, rather than relying only on a completion summary. A preview may be useful while still being unsuitable for a time-critical or regulated process.

09 / The story

Claude Duet versus chat and coding agents

These categories describe common workflows, not identical behavior across all products. Some chat applications and coding agents also support background work or integrations. Compare the actual setup and permissions rather than assuming the label tells you everything.

010 / The story

Who can access Claude Duet, and does it cost extra?

As checked October 22, the research preview requires a Claude Pro or Max plan, a US Claude account and an invitation from the waitlist. Invitations arrive in waves, so signing up does not establish an access date. Team and Enterprise are not included in this initial preview. The product is a standalone iOS or Android app and can also be opened on the web at claude.ai. [1] [3]

During the preview, the Help Center says Duet is included with Pro and Max at no extra cost. It shares existing Claude usage, including five-hour session and weekly limits. Morning briefs and background checks count; larger tasks may consume more usage. “Included” does not mean unlimited or cost-free in every future version. Eligibility and terms may change, so check the official pages before upgrading. [1] [3]

The opportunity is less coordination work: prepared drafts, surfaced issues and routine follow-through. The trade-off is delegating access to a system that can make mistakes. Would you trust it with a narrow inbox-and-calendar task—and what would you insist on reviewing yourself?