On October 21, 2026, Anthropic introduced Claude Duet, a proactive AI agent that works in the background, connects to a user’s apps and flags what needs attention. It is a research preview with a waitlist, limited to Pro and Max subscribers in the United States.
What happened
According to MacRumors (October 21, 2026), Duet is built for professionals and small business owners with a lot to keep track of. It connects to the apps on a Claude account, spots open items and completes tasks, including in the cloud without local access to the user’s computer: it has a virtual machine with its own browser. It is positioned against Meta’s Muse and OpenAI’s Dots, which also reach out before the user asks.
Key facts
- It runs on Opus 5.5 and draws on the memory, skills and connected apps of the Claude account, all in one ongoing thread.
- Each morning it sends a short brief on important items, and it checks connected apps every few hours, alerting the user when something needs attention.
- It asks for approval before actions such as sending a message, making a payment, signing a document or deleting a file.
- It does not see passwords or logins when given account access, and a dedicated classifier and system prompt screen messages for sensitive information leaks.
- It uses existing plan limits: its background checks count toward them.
- Availability: waitlist, Pro and Max in the US only, on the web and in a standalone iOS app.
Context
Another roundup this week, from LeverageFirst, also says Anthropic plans to make Claude’s «auto mode» the default for Pro, Max and Team plans starting the week of October 26, with automated checks that block risky actions; MacRumors does not mention that point, so we don’t treat it as confirmed. What is confirmed is the trend: the three big players, Meta, OpenAI and Anthropic, now offer agents that act before being asked.
Strategic takeaway — Studio 24-7
A proactive agent is only as useful as the business it plugs into is organized. If processes, customer data and approval rules aren’t defined, the agent just speeds up the mess. Duet’s human approval for payments, signatures and messages is a good design model: automate what is repeatable, keep control over what is irreversible.
- Map which tasks in your operation are repeatable and which need human judgment.
- Decide now which actions an agent should never execute without approval.
- Clean up the tools that would stay connected: an agent performs according to what it can see.
Systems before improvisation: the agent is the last layer of a system, not the first step.
Sources
- MacRumors — Anthropic Claude Duet (October 21, 2026)
- LeverageFirst — AI News This Week (October 24, 2026)