Anthropic has opened a research preview of Claude Duet, an AI agent that runs on its own cloud computer and messages the user when it decides something needs attention.
The preview is limited to Claude Pro and Max subscribers in the United States, who join a waitlist. It costs nothing extra during the preview and is not yet available to Team or Enterprise customers. Duet runs on Claude Opus 5.5.
An assistant that starts the conversation
Most AI tools wait for a prompt. Duet works in the background, reading connected apps and producing a morning brief. It then sends follow-up messages during the day when it finds a deadline, an unanswered email or a task that can be completed.
In Anthropic's examples, the agent can produce research briefs, financial models, sales forecasts, slide decks, dashboards, prototypes and small web apps without being asked to start. The company describes the target users as professionals and owners of small businesses who need help with ongoing work rather than one-off questions.
Approval before anything irreversible
Anthropic said Duet asks for approval before sending messages, making payments, signing documents or deleting files, and that the approval request appears with the exact action it plans to take. Users can adjust how often it asks.
The agent's credentials are handled the way Anthropic handles its Managed Agents product, so Duet does not see passwords or tokens for connected services. A classifier reviews actions and flags those that need a human decision.
Duet uses the same usage limits as the rest of a subscriber's plan, and background check-ins count against them. Anthropic said it expects users to hit limits faster than they do with chat.
A crowded field
Anthropic is not first. Meta released Muse on 8 September, OpenAI launched dots on 29 September and xAI shipped Grok Bot in August. All three pitch a persistent assistant that remembers context and acts without being prompted, and all three have run into questions about how much autonomy users actually want.
Early testers have reported that proactive agents can be useful for scheduling and research but noisy when they misjudge what matters. Anthropic's answer is a visible approval step and a narrow set of connected apps at launch. The company said it will use the preview to learn which notifications people keep and which they turn off.