Anthropic Opens Claude Duet as a Proactive AI Agent
Anthropic is testing Claude Duet, an agent that monitors priorities, completes tasks and asks for approval before sensitive actions.
Anthropic has opened a research preview of Claude Duet, a proactive agent designed to work across a user’s digital tasks rather than wait for individual prompts. The service is available through a US waitlist for Pro and Max subscribers.
Anthropic describes Duet as an application built around Claude Opus 5.5. It can identify work that may need attention, complete tasks in the background and return to the user when a decision or approval is required. The company says projects can span repositories from GitHub, GitLab and other providers, with findings delivered through a new API and MCP connector.
The product’s design reflects a shift from conversational assistants to systems that maintain an operating view of a user’s work. That creates a more useful workflow for software teams and knowledge workers: the agent can collect context, inspect connected sources and carry out a sequence of actions instead of producing a single answer. It also creates a harder safety problem, because mistakes can occur between the initial instruction and the eventual action.
Anthropic says Duet does not directly see users’ passwords and uses a classifier to guard against attempts to share sensitive information. The company also acknowledges that the system can misjudge what is urgent. Those limits matter because proactive behavior turns ambiguity into an operational decision: waiting may be wrong, but acting too early may be worse.
The early research-preview status and US-only availability make this a controlled experiment rather than a mass-market launch. Its significance lies in the product direction. Anthropic is testing whether a model can become a persistent work partner while remaining interruptible, permission-aware and bounded by security controls. If the approach works, agent products will compete less on chat quality and more on judgment under incomplete information.
Uncle Cat take
Claude Duet’s defining risk is not access to repositories; it is deciding which unfinished task deserves interruption, a judgment Anthropic admits remains imperfect.