Claude Duet: Anthropic's Proactive Agent Runs on Opus 5.5 with Approval Gates

Anthropic launched Claude Duet, a research preview of what it calls a proactive agent. It runs tasks on its own computer in the cloud, uses connected apps, and messages you when it finishes or needs a decision. It is powered by Opus 5.5 and available on Pro and Max plans in the US via waitlist.

What it actually does

Duet is designed around a hand-off model: you send a message, it works asynchronously, and it returns when the task is done or blocked. The product page shows a morning-brief workflow where the agent surfaces items that need attention and then goes quiet. It also demonstrates multi-step execution: building a dashboard from POS, Shopify, CRM, support inbox, payroll, and bank data, then refreshing it daily at 5:30 AM.

• Model: Opus 5.5
• Availability: Pro and Max plans, US only, waitlist for research preview
• Connectors: more than 6,000 connectors and plugins
• Memory: memory, skills, and apps from the Claude app carry over from the first message
• Approval gates: Duet asks before sensitive operations like sending messages or paying
• Visibility: you can see what it has done and is doing, and steer it mid-task

Safety and execution model

The agent runs in a cloud computer with approvals, logins, and safety checks scoped to that environment, so sensitive details stay private. The examples on the page show draft-and-send flows: Duet prepares an email to a VC partner with a data room link, then waits for an explicit send action. The same pattern applies to payments — a county health permit fee of $640 is presented with an expiry countdown and an approve/not-now choice.

Example tasks from the source

The launch page walks through a few concrete workflows rather than abstract capabilities:

• Building an "opening checklist" for a new location from two prior openings (41 items, 33 done or on track, 7 watched, 1 requiring approval)
• Rebuilding a Series A deck with Q3 data (12 slides) and drafting partner-meeting prep from public bios and prior deals
• Investigating a churn spike (East Bay, 2.6% to 3.4% after a shipping price change) and proposing a 2-week free-shipping test
• Pulling missing data-room documents from connected sources (23 of 26 to 25 of 26 after approval)

Who it's for

Operators and small-team founders who want an agent to run multi-step work across business apps — not just answer questions in a chat window — with explicit approval before anything irreversible happens.