1. News
  2. AI models & agents
  3. Anthropic · AI agents
Anthropic · AI agents3 min read

Claude Duet, an AI that works while you're away, asks before it acts

Claude Duet, a research preview agent from Anthropic's Claude team, can work in your accounts while you're away. It shows what it plans to do first.

By the Dashline deskSource: Claude engineering
Claude Duet, an AI that works while you're away, asks before it acts
Photo: Alphaaloy, CC BY-SA 4.0

Claude Duet is an AI agent that keeps working after you close the app, and it is built to stop and ask before it does anything risky. The Claude team published a long post on its safety, privacy and security. It is still a research preview, so it is an early version open to testing.

What it does

Claude Duet uses a cloud-based computer to do work for you. It can use the apps and services you have connected to Claude. It can read your inbox, draft replies, book things and run routines while you are away. The team says it uses Claude Duet on its own accounts for work every day.

How it asks first

Before it sends a potentially sensitive message, pays for something or deletes anything, Claude Duet shows you the action, the destination and the content, including every recipient. Then it waits for your approval. A separate automated check also tests whether each action matches what you wanted. That check runs outside the model doing the work, so instructions hidden in a web page cannot sway it.

Some actions are blocked even if you approve them: trading securities, moving money between bank accounts and accessing someone else's account. Paying, sending money, deleting an order and creating an account always need a fresh approval. A "don't ask me again" rule does not cover them.

Money and passwords

For purchases, Claude Duet uses a one-time card from a wallet like Link from Stripe. The card works only for that merchant and amount. You see the item and total and approve it in your Link app. Your real card number is not shared with Claude Duet or the merchant.

When a site asks for a password, you type it into a form in the Claude app. It is encrypted with a single-use key and passed to the browser without being opened. The model never sees it. Sign-in with 1Password is expected soon.

Privacy and memory

A separate safety check looks at each outgoing message for sensitive information. If a message is blocked, Claude Duet can rewrite it without that information and ask you to approve the new version. The team admits the checks will sometimes miss something or flag a harmless message.

  • Claude Duet keeps your history and memory until you delete or reset them.
  • Deleting your account wipes everything it has stored.
  • Disconnecting an app stops new access, but what it already learned stays until you reset memory.
  • Memory from your earlier Claude chats is visible to it.

What to do

Before turning on an agent like this, check which apps it can reach, what it can send and what it can pay for. Read each approval request closely. The team notes that too many requests can teach people to tap through, and says it will keep adjusting how often it asks.

The takeaway

Before letting an AI act for you, check what it can see, send and pay for, and read each approval request instead of tapping through.

More articles

You should really read these.

Next step

Let’s talk