The news
Anthropic introduced Claude Duet on October 21. The agent runs outside the main Claude interface and contacts users with updates or completed work rather than waiting for prompts. It targets professionals and small business owners who manage multiple apps and cannot track every notification.
Context
Prior agents from the same category, including Meta's Muse and OpenAI's Dots, already operate on similar principles. Those tools also maintain their own virtual machines and browsers to act on behalf of the user. Duet follows the same pattern but ships with its own daily briefing cadence and explicit focus on administrative load. It is built on Opus 5.5 and does not require the user to keep a local computer running.
The announcement positions Duet as a distinct product line rather than an add-on inside the existing Claude chat. Anthropic describes it as a personal agent meant to reduce the burden of monitoring scattered accounts. In practice this means the service keeps its own session state and cloud environment active at all times. Users connect their apps once, then receive periodic outreach instead of opening the main Claude window to issue new instructions.
Details
Duet connects to the user's apps and sends a short morning brief that lists items requiring attention. It then polls those apps every few hours. When it detects something that needs action, it either alerts the user or attempts to finish the task itself inside its cloud virtual machine. The agent can complete simple administrative steps without further input once the initial connection is made. Anthropic states that the service is intended for people who already carry too many open threads across tools and are therefore most likely to benefit from an always-on monitor.
The product is positioned as a personal agent rather than an extension of the core Claude chat experience. It maintains its own session state and browser environment so tasks can proceed even when the user is offline. No additional hardware or local software installation is required beyond the initial account link. The company has not released pricing or exact availability dates beyond the announcement.
MacRumors reporting notes that Duet reaches out before the user asks, mirroring the approach taken by Muse and Dots. The agent uses its dedicated virtual machine and browser to carry out steps inside connected services. Thurrott coverage adds that the delivery happens outside the standard Claude user experience, keeping the proactive behavior separate from on-demand conversations. Both sources confirm the morning brief plus periodic checks as the primary rhythm of interaction.
The underlying model is Opus 5.5. Anthropic has not disclosed how many apps the agent can link to at launch or what specific connectors are supported. The emphasis remains on administrative work that repeats across email, calendars, and project tools. Once permissions are granted, the agent can act inside its isolated environment without requiring the user to stay signed in on a personal device.
Why it matters
Duet shifts the interaction model from on-demand chat to scheduled outreach. Users who adopt it will receive fewer raw notifications and more pre-filtered summaries or completed actions. For small teams already stretched across email, calendars, and project tools, the change reduces the daily cost of context switching. At the same time, the agent still depends on the quality of its underlying model and the reliability of the app connections it is given. If those links are incomplete or the model misreads priority, the briefings and automated steps lose value quickly. The release therefore tests whether professionals will grant an external agent steady access to their accounts in exchange for fewer manual checks.
The move also places Anthropic directly alongside Meta and OpenAI in the emerging category of always-running agents. Each company now offers a tool that operates in the background rather than waiting inside a chat window. Differences will likely appear in the breadth of supported apps, the accuracy of priority judgments, and the willingness of users to accept automated actions without review. Early adopters will reveal whether the morning brief plus cloud execution model reduces real workload or simply adds another layer of messages to review.
---
Sources:
No comments yet