Anthropic unveiled Claude Duet on Wednesday, joining a rapidly crowded market for AI agents that can take actions rather than merely answer prompts. The company’s bet is that professionals and small-business owners need a vigilant work assistant more than another tool for booking dinner reservations or travel.¹

Duet is being released as a research preview with a waitlist, limiting its immediate reach. But Anthropic is pitching a more persistent kind of software: an agent built on its Opus 5.5 model that monitors connected apps, prepares a morning brief and flags matters needing attention. Addy Osmani described it as an app that “keeps going when you step away,” with access to 6,000 connectors and plugins.Addy Osmani (addyosmani)²

That framing puts Anthropic in direct competition with OpenAI’s Dots, Meta’s Muse, Google’s Gemini Spark and newer agent start-ups. The difference, Anthropic says, is focus. While OpenAI and Meta have emphasized daily personal tasks, Duet was built for work and is less capable at personal errands.³

The company’s intended workflow ranges from scheduling meetings to building dashboards, forecasts and web apps. Yet the more consequential the task, the more obvious the trust problem becomes. Duet uses a virtual machine and browser to carry out work, but is meant to seek approval before major moves such as sending payments or deleting information.⁴

Anthropic is also conceding that the agent is unfinished: it may misread a thread or misjudge urgency, and it remains weaker at errands such as booking tickets. That makes Duet’s launch less a declaration that autonomous agents have arrived than a carefully bounded wager that workers will tolerate an AI colleague—provided it knows when to ask first.⁴