Anthropic 发布主动式个人智能体应用 Claude Duet，对标 Grok Bot、Meta Muse 与 OpenAI Dots

Anthropic has answered Grok Bot, Meta's Muse, and OpenAI's Dots with a research preview of Claude Duet. The standalone app runs Claude as a proactive agent that keeps working on tasks while the user is away and asks for approval before taking sensitive steps.

Anthropic is pitching Claude Duet to professionals and small business owners who have "more on their plates than they can track." The app runs on iOS, Android, and the web at claude.ai and is built on Claude Opus 5.5. Pro and Max subscribers in the US can get access through a waitlist. Team and Enterprise plans aren't included yet.

Anthropic says the big difference from the regular Claude app is which way things flow. In the Claude app, users open a chat. Duet comes to them instead. Every morning, the agent sends a short summary of the most important items from connected apps. After that, it only reaches out when something needs attention. It pulls memory, skills, and app connections from the user's existing Claude account.

Users assign tasks by text or voice message in a single ongoing thread. According to Anthropic, that covers everything from scheduling to dashboards and financial forecasts to prototypes and web apps. The agent works on a virtual machine, its own cloud computer with a browser, so the user's laptop or phone doesn't need to stay on.

Duet connects to services like Gmail, Slack, Stripe, Docusign, and Shopify through more than 6,000 connectors and plugins from the Claude Marketplace. That lets it track down unpaid invoices and draft payment reminders, or get payroll ready. For purchases, it uses Stripe Link with a virtual card, so the user's real card number never gets shared.

Several competitors have launched similar products in recent months. SpaceXAI opened the Grok Bot beta in August. In September, Meta followed with Muse and OpenAI with its Dots. All of them are betting on personal agents that handle tasks independently and act on their own initiative.

Users keep the final say on anything risky

Anthropic says Duet can only read, draft, and prepare on its own. Before it sends a message, makes a payment, signs a document, or deletes anything, it shows exactly what it plans to do and waits for the user's approval. Trading securities and moving money between accounts are off the table, even with approval.

Technical safeguards add another layer. The company says the agent never sees passwords or login tokens directly, because credential handling follows the design of Claude Managed Agents. To fend off hidden instructions, known as prompt injections, Anthropic uses defenses from its research on browser use. Red teams and an outside firm also ran attack tests to probe Duet for weak spots.

Anthropic also built a new classifier for Duet. This screening system works with special safety instructions in the system prompt to check every message for leaking sensitive data. Together, they're meant to refuse requests to share credit card numbers or ID details. Health or personal finance information only gets through when the user explicitly says so.

Anthropic admits Duet still misses things

Anthropic concedes that the early version can misread a message thread, misjudge urgency, and miss things that arrive between two checks. It's also weaker at personal errands like booking tickets, since its core features are built for work.

Duet only checks connected accounts every few hours, and Anthropic says that comes down to usage. Constant checking would burn through subscription limits much faster. Duet shares those limits with other Claude products, and the background checks count against them too.

Anthropic plans to use tester feedback to tune when the agent should reach out. How Duet fits with the Claude app is still an open question. For now, it runs separately, and early users' experience will shape how the two work together.

AI News Without the Hype – Curated by Humans

Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.