Anthropic Launches Claude Duet Agent With Approval Gates for Sensitive Actions
The new proactive agent for Pro and Max subscribers requires explicit user consent before sending messages, making payments, or deleting files.
Key points
- Claude Duet operates in a cloud-based virtual machine with its own browser, allowing background task execution without local device access.
- The agent integrates with over 6,000 applications, including Gmail, Slack, and Shopify, leveraging existing memory and skills from user accounts.
- Users must approve high-risk actions such as document signing or fund transfers, while the system blocks securities trading and inter-account money moves.
Anthropic introduced Claude Duet, a proactive artificial intelligence agent designed for professionals and small business owners, entering a research preview phase on 21 October. The product targets users who need assistance tracking tasks across multiple platforms, offering a background assistant that operates independently of the user’s immediate attention or device status.
The agent runs on the Opus 5.5 model and utilizes a dedicated virtual machine equipped with a browser, enabling it to complete tasks in the cloud. According to MacRumors, this architecture allows Duet to function without requiring local access to a user’s computer, distinguishing it from tools that rely on device-side processing. It connects to existing apps to suggest items needing attention and can execute tasks ranging from simple administrative duties to building web applications.
Context and Background
Anthropic positions Claude Duet as a competitive response to similar agents from other major technology firms, specifically Meta’s Muse and OpenAI’s Dots. As reported by The Next Web, these competing products also feature proactive capabilities that reach out to users before they initiate requests. Duet differentiates itself by integrating deeply with the existing Claude ecosystem, inheriting memory, skills, and app connections from a user’s current account without requiring extensive setup.
The agent sends a morning brief summarizing important items from connected applications and checks those apps every few hours for new alerts. SiliconANGLE notes that Duet is built upon Anthropic’s framework for trustworthy agents, which also guides systems like Claude Code. This framework includes an automatic classifier that scrutinizes messages for sensitive information leaks and verifies actions with the user before proceeding.
Who Is Affected
The research preview is currently limited to Pro and Max subscribers in the United States, with access granted through a waitlist. Users can interact with Duet via the web, a standalone iOS app, or Android, according to The Next Web. The service is included in existing subscriptions at no additional cost, though the background checks and task execution count toward the user’s standard plan limits.
Anthropic has outlined specific limitations for this early version, noting that Duet may misjudge message urgency or struggle with personal errands like booking tickets. Early reviewer Dan Shipper of Every found the email drafts weaker than those from competing agents and noted that voice features are currently limited to transcription rather than a full voice mode. Users cannot yet observe the agent’s browser activity in real time.
What should you do?
Security professionals and power users should monitor the research preview for changes in how the agent handles sensitive data and approval workflows. As the agent gains access to numerous connected applications, verifying the scope of permissions granted to each integration is critical. Users should review the specific actions that require explicit approval versus those that are automatically blocked, such as securities trading.
What to do and how to stay safe: Anthropic
- Audit the list of connected applications to ensure Duet only has access to necessary services, reducing the potential attack surface if the agent is compromised.
- Review the approval settings to confirm that high-risk actions, such as payments and document signing, always require explicit user consent before execution.
- Monitor usage limits closely, as background tasks performed by Duet count toward your existing subscription caps, which could impact availability for other Claude features.
- Keep an eye on updates to the agent’s safety classifiers and permission models, as Anthropic continues to refine the system based on early user feedback.
Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
Does Claude Duet require a new subscription?
No, it is included for existing Pro and Max subscribers in the US, but it uses your current plan’s usage limits.
Can Duet make financial transactions without my permission?
No, it requires user approval for payments and blocks securities trading and transfers between accounts entirely.
Is Claude Duet available globally?
Currently, the research preview is limited to subscribers in the United States.
