AI Trend Research
EN

← AI glossary

AI glossary · Basic · Agents · Language models

Claude Duet

Claude Duet is a personal AI agent that Anthropic released as a research preview. It handles tasks assigned via message on a cloud computer, keeps working while the user is away and reaches out first when a decision is needed.

About this term
Level
Basic · Technical terms you often meet in AI news
Added
2026-10-22
In plain words It is similar to an assistant who lays out what you need to handle when you arrive at work, works on assigned tasks independently and always asks before spending money or sending an email. Unlike a human assistant, however, it may misjudge what is urgent or misread messages.

What's different

In the existing Claude app, the user opens a chat and works alongside it. Duet works the other way around: the agent reaches out first and carries out tasks handed to it via message on its own. Anthropic said the service is aimed at professionals and small business owners with a lot of work and decisions to handle. There is no need to write prompts or configure settings; you assign work by text or voice, as you would to a colleague. Everything from short replies to week-long projects continues in a single conversation thread.

Existing Claude app vs. Claude Duet

Existing Claude appWorking together

  • User opens a chat to start
  • Works together through conversation

Claude DuetDelegate and wait

  • Agent reaches out first
  • Works alone on a cloud computer
  • Asks for approval at key steps

공통 Uses the same account's memory, skills and connected apps

Compares who starts the conversation and where the work happens.

How it works

Duet works on a cloud virtual machine with a browser, using connected apps. As a result, the user's laptop or phone does not need to stay on. Every morning it sends a briefing that gathers important items from email, calendar and connected apps, and after that it checks in every few hours and reaches out when something needs attention. Anthropic says it was designed this way because constant real-time checking would use up usage limits quickly, and users can set what they are notified about and how often. It uses the memory, skills and connected apps of the existing Claude account from the first message, and Anthropic said it can integrate with more than 6,000 connectors and plugins.

Duet's daily loop

  1. 1. Morning briefing Summarizes key items
  2. 2. Assign work By text or voice
  3. 3. Works in the cloud Uses connected apps
  4. 4. Approval request Before sending, paying, deleting
  5. 5. Periodic check-ins Every few hours
After the morning briefing, it handles assigned tasks, checks in every few hours and calls on the user only when needed.

Why it matters

It shows a shift in AI agents from tools that respond only when called by the user toward colleagues that take care of things on their own and reach out first. According to the announcement, it can handle tasks ranging from scheduling and drafting replies to checking unpaid invoices, preparing payroll and creating presentations, dashboards and web apps. Media outlets reported it as a product competing with Meta's Muse and OpenAI's Dots. As of October 2026, it is in a research preview based on Opus 5.5, and Pro and Max subscribers in the US can sign up for the waitlist.

Things to know (limitations)

It reads materials and creates drafts on its own, but for hard-to-reverse actions such as sending messages, making payments, signing documents or deleting items, it first shows what it plans to do and waits for user approval. Some actions, such as securities trades and transfers between accounts, it will not perform even with approval. Anthropic said the agent is designed not to see passwords or login tokens directly, and that outgoing messages are checked by a dedicated classifier for sensitive information. Anthropic itself said that, as an early version, it may misread conversation context or misjudge urgency, and that it is still weak at personal errands such as booking tickets. Team and Enterprise plans are not yet included, and background checks also draw on the usage limits of existing subscriptions.

Example

The demo featured in this site's news item 'Anthropic, 사용자가 없을 때도 일하는 개인용 에이전트 Claude Duet 리서치 프리뷰 공개' (Anthropic releases research preview of Claude Duet, a personal agent that works even when the user is away) shows how it can be used. When the owner of a café chain is preparing to open a new location, Duet builds a checklist from records of previous openings, prepares the health permit application and fee payment, and then sends a message with an 'Approve' button. Once the owner approves, it completes the payment, requests an inspection date and keeps tracking delays in equipment installation. When preparing for fundraising, it updates the pitch deck with the latest quarterly figures and drafts emails to investors, but only sends them after the owner approves.

Where it came up on this site

Related terms

AI agent Always-on agent Human-in-the-loop Computer use Agent memory Skills

References