Claude Duet: what Anthropic's proactive background agent means for agent builders

Anthropic has launched Claude Duet, a proactive agent that keeps working when you aren't prompting it. It is a research preview, gated by a waitlist, and per MacRumors and other launch coverage it opens first to US Pro and Max subscribers. Team and Enterprise plans are not included yet. Coverage frames it as Anthropic's answer to the background agents from Meta and OpenAI.

What Duet actually does

Duet runs on Claude Opus 5.5 and draws on the memory and skills already attached to a user's Claude account. It runs as a standalone app on iOS and Android and on the Claude website, separate from the main Claude app during the preview.

• Morning brief: a short summary of the most important items across connected apps such as email and calendar.
• Periodic check-ins: after the brief it checks in every few hours and messages you only when something needs you.
• Approval gates: it asks before sending a message, making a payment, signing a document or deleting a file.
• Shared quota: background checks count toward the plan's usage limits, the same as regular chats.

Anthropic has said Duet can still misjudge what is urgent. Coverage also notes that personal errands are less developed and that users cannot yet watch its browser.

Why builders should care

The consumer product is the least interesting part for our readers. The architecture is the useful part, because it shows the shape the market is converging on for agents that act without being asked.

Proactivity is a triage problem

An agent that runs on a schedule has to decide when to interrupt. Duet's pattern is a digest by default and a push message by exception. Anthropic's own admission that it can misjudge urgency is a reminder that this is the hard part. A false positive wastes attention, and a false negative defeats the point of having the agent.

Approval gates mark the irreversible actions

The listed actions are sending, paying, signing and deleting. These are the side effects you cannot take back. In our own agent work, the useful design question is which actions are reversible. Read-only work can run unattended. Anything that moves money or data out of your control should wait for a human.

Always-on agents have a running cost

Because background checks draw on the same quota as chats, an idle-looking agent still spends budget. If you build something similar, budget for polling frequency and set a ceiling on how often the agent wakes.

What we'd watch next

It is early. The launch is a waitlist preview, and we haven't seen Anthropic publish measured reliability for its proactivity judgment. We would treat any claim about how well it triages as unproven until there is independent evidence. Pay attention to how the approval gates behave under real, messy workloads, and to whether the pattern reaches Team and Enterprise plans, where audit and permission requirements get stricter.

For anyone shipping agentic tooling, the takeaway is that the product category now has a baseline: scheduled wake-ups, quiet-by-default messaging and explicit consent for irreversible actions. Users will expect those behaviors from your agent as well.