The lean read

What happened?

Anthropic opened a waitlisted research preview of Claude Duet, a proactive agent designed around Claude Opus 5.5.

Why does it matter?

Background agents must decide what deserves attention and when to act. Anthropic’s own limits and approval safeguards are therefore central to evaluating the product.

Follow the evidence ↗

What happened?

Anthropic opened a research preview of Claude Duet on October 21. The product is designed to act as a proactive assistant: it can bring users a short morning brief, monitor connected apps, continue work while they are away and ask for confirmation before acting on their behalf. Anthropic says the preview is available through a waitlist to U.S. Pro and Max subscribers, as a standalone iOS and Android app and through claude.ai.

Anthropic says Duet is designed around Claude Opus 5.5, the model it introduced on September 22. The company describes tasks ranging from scheduling to building dashboards, forecasts and web apps, but these are product capabilities claimed in an early research preview, not an independent performance evaluation.

Source: Anthropic ↗ · Source: Anthropic ↗

The limits are part of the launch

Anthropic says Duet may still misread a thread or misjudge what is urgent. Personal errands are weaker than work tasks, the product cannot yet show users its browser, and background checks count against plan usage limits. The preview also lives in a separate app from the main Claude app.

The company says Duet is built so it does not directly see passwords or sign-in tokens, and that an auto-mode classifier helps determine when actions need approval. Those safeguards are part of Anthropic’s design account; they are not proof that an always-on agent will make every decision correctly.

Source: Anthropic ↗

Our take

Why does it matter?

Our analysis: Claude Duet shifts the assistant model from answering when prompted to deciding when a user should be interrupted. That can reduce the effort of managing a busy set of apps, but it also makes prioritization and restraint central product behaviors. A useful agent must know not only how to complete a task, but when not to create another one.

The waitlist and research-preview label set a realistic boundary around the announcement. The meaningful test will be whether users can understand why Duet surfaced something, what it did in the background and where approval remains necessary.

Read the sources

  1. Introducing Claude Duet ↗
    Anthropic · October 21, 2026
  2. Introducing Claude Opus 5.5 ↗
    Anthropic · September 22, 2026
  3. Claude Duet: Anthropic’s Agent Runs on Opus 5.5 ↗
    Shattered.io · October 21, 2026