Anthropic has launched Claude Duet, a new work-oriented AI agent designed to handle ongoing professional tasks either on its own or through integrations with external applications.
The company describes it as capable of continuing work independently even after a user steps away, while still requiring explicit approval before performing sensitive actions such as sending messages or initiating payments.
Certain madia outlets have situated the release within an accelerating competition among major AI firms, specifically noting parallels with offerings from Meta and OpenAI.
Anthropic CTO Rahul Patil addressed related concerns directly on LinkedIn, observing that granting an AI system access to tools like email, Slack, and calendars can understandably feel unsettling.
For the time being, Claude Duet remains available only as a waitlisted research preview limited to subscribers on Anthropic’s Pro and Max plans.Official details from Anthropic frame Claude Duet as a proactive agent built around the company’s Claude Opus 5.5 model and aimed primarily at professionals and small-business owners who manage more responsibilities than they can easily track.
It arrives as a standalone application for iOS and Android as well as through the claude.ai website.
Once connected, it can draw on existing memory, skills, and app linkages already associated with a user’s Claude account.
The agent is intended to surface priorities rather than wait for constant prompts: it can deliver a concise morning summary of important items drawn from linked services, monitor those services during the day, and reach out only when something appears to require attention.
Users may also assign tasks ranging from straightforward administrative work, such as scheduling meetings, to more involved efforts like assembling dashboards, preparing financial or sales forecasts, or developing prototypes and web applications.
A core design principle is that consequential decisions remain with the human user.
Claude Duet can read information, draft content, and prepare materials in the background on its own cloud-based virtual machine equipped with a browser.
Before it sends a message, completes a payment, signs a document, or deletes a file, however, it is designed to present the proposed action and pause for confirmation.
Certain activities, including trading securities or transferring money between accounts, are restricted even with approval.
The system also incorporates safeguards intended to limit the exposure of highly sensitive details such as credit-card numbers or government identification.
These features place Claude Duet in the same emerging category as other recent agent products that operate persistently in the cloud, monitor connected apps, and act with varying degrees of autonomy.
Anthropic has acknowledged that the current research-preview version is still early and may occasionally misinterpret context or misjudge urgency; it is also described as less capable with purely personal errands such as booking tickets.
Access during the preview is restricted to eligible Pro and Max users in the United States, shares the same usage limits as other Claude features, and is being rolled out through a waitlist.
The combination of proactive monitoring, background execution, and explicit permission gates reflects an attempt to balance utility with caution as AI systems gain deeper access to workplace tools. Early commentary from Anthropic’s leadership and coverage from technology outlets both highlight that trust—particularly around data access and the potential for mistakes—will likely shape how widely such agents are adopted.
