Claude Duet Is Anthropic's Proactive Agent: It Comes to You Instead of Waiting

Anthropic launched Claude Duet on October 21, and the interaction model is the headline: instead of a chatbot that waits for prompts, Duet is a proactive agent that initiates contact. Powered by Opus 5.5, it sends a daily morning brief of what needs your attention, drafts next steps, then goes quiet until something requires you. Large tasks, from meeting scheduling to building dashboards, decks, and documents, run in a cloud computer while you are away, with the agent pinging you when work is done or when an action needs approval. The launch puts Anthropic in direct competition with Meta’s Muse in the consumer agent race, and coverage from TechCrunch, VentureBeat, The Verge, and the New York Times framed it exactly that way.

How the proactive loop works

Duet’s design answers the question every agent demo dodges: what does the agent do between tasks? The answer is a scheduled attention cycle. A daily brief opens the thread, the agent proposes next steps, and accepted work executes on Anthropic’s cloud infrastructure, with the demo showing a “Marlo Pulse” dashboard that refreshes at 5:30 AM before the brief lands. The agent shares your existing Claude account’s memory, skills, and app integrations in one thread, works with 6,000-plus connectors and plugins, and handles tasks like chasing unpaid invoices, preparing payroll, and contract prep. The stated safety model is approval-gated for sensitive actions such as sending messages or payments, with approvals, logins, and safety checks executing inside the user’s own cloud computer rather than Anthropic’s, and a full view of everything the agent has done or is doing at all times.

Availability, and what the limits reveal

Access is a research preview with a waitlist, restricted to Claude Pro and Max subscribers in the US, with Team and Enterprise plans explicitly excluded for now. Usage draws from the existing plan’s limits, with larger tasks consuming more. Anthropic’s own limitations section is more candid than most launch pages: Duet may misread threads, misjudge urgency, or miss items that arrive between check-ins. That honesty maps onto the real engineering challenge of proactive agents, which is not capability but calibration: an assistant that interrupts too often gets muted, and one that stays quiet through something important loses trust permanently. This is the design problem Anthropic has been circling all year, from desktop control in the Claude apps to managed agents in Claude Code, now extended to the consumer morning-brief form factor.

The competitive frame, and what to watch

The market context explains the timing. Meta reported this week that Muse hit 10 million US downloads in 40 days, only the twelfth app ever to do that, with Zuckerberg claiming on the earnings call that more than a quarter of Muse users make purchases averaging over $200. Anthropic has no social graph and no ad business to subsidize distribution; what it has is the perception of professional-grade trust, and Duet’s whole posture, approval gates, private cloud execution, no engagement-maximizing feed, is built for the user Muse is not designed for. What to watch: whether the waitlist opens beyond the US and Pro tiers before the end of the year, how Duet’s task-completion economics survive contact with Pro plan limits (an agent that runs dashboards overnight is expensive at $20 a month), and whether Meta answers with proactivity in Muse, because the two products are currently betting on opposite ends of the attention spectrum. The agent race just moved from who can answer best to who can speak first, appropriately, without being annoying.