Anthropic Launches Claude Duet, an AI Agent That Watches Your Inbox So You Don't Have To

Anthropic has entered the crowded race to build an AI assistant that actually runs your workday instead of just answering questions about it. The company announced Claude Duet on Tuesday, a new agent aimed at professionals and small business owners that monitors your connected apps, flags what needs attention, and takes on tasks ranging from scheduling meetings to building financial forecasts.

Duet arrives as a "research preview," Anthropic's term for an early-access product still being refined based on real-world use. It's available now as a standalone app on iOS and Android, as well as through Claude on the web, though access is gated: you'll need a Pro or Max subscription in the US, and interested users currently have to join a waitlist.

The agent is built on Claude Opus 5.5, Anthropic's latest model, and the pitch is less "chatbot" and more "proactive coworker." According to Anthropic's own description, Duet sends a short morning brief covering the most important items across the apps it's connected to, then keeps watch throughout the day, interrupting you only when something actually needs a decision. Beyond monitoring, it can take on assigned work: simple administrative tasks like scheduling, or heavier lifts like assembling multi-tool dashboards, drafting sales forecasts, prototyping, and building web apps.

To do any of that, Duet taps into the broader Claude ecosystem, pulling memory, skills, and connected-app access from the main product. It also runs on a virtual machine equipped with a browser, which lets it navigate the web and complete tasks the way a person would rather than relying solely on API integrations.

Guardrails before it acts

Because an agent with this much reach is also an agent that can make costly mistakes, Anthropic built in a checkpoint system. Before any consequential action — sending a payment, deleting a file, anything with real stakes — Duet pauses and messages the user for explicit approval rather than acting autonomously.

Anthropic also addressed the risk that comes with an AI touching sensitive personal and financial information. The company says Duet includes a new dedicated classifier paired with a safety system prompt specifically designed to reject requests involving what it calls the strictest tier of sensitive data, citing credit card numbers as an example. In plain terms: certain categories of information are walled off from being shared or acted on, regardless of what's asked.

Anthropic isn't overselling it

Notably, Anthropic is upfront about Duet's limitations rather than burying them. The company describes it as an "early version" that "still might misread a thread or miscalibrate what's urgent" — an acknowledgment that the core premise of the product, correctly judging what deserves your attention, is also its hardest problem to get right. Anthropic also flagged that Duet isn't yet as capable with personal errands, like booking tickets, as it is with professional and business tasks.

That combination of ambition and caveats fits the research-preview framing: this is a product meant to improve through real usage and feedback, not a finished tool shipped with full confidence.

The agent wars are heating up

Claude Duet lands squarely in the same category as Meta's Muse and OpenAI's Dots — AI agents designed to sit inside your digital life, watch your apps, and act on your behalf rather than wait for you to open a chat window and type a prompt. That three-way competition signals where the major AI labs think the next phase of the product fight is happening: not in raw model benchmarks, but in who can build the most trustworthy autonomous layer on top of those models.

For Anthropic specifically, Duet is a bet that its existing base of professional and business users, who already rely on Claude for coding and document work, will want an agent that extends that relationship into ongoing, hands-off monitoring. The approval-gating and sensitive-data classifier suggest Anthropic is trying to thread a difficult needle: give the agent enough autonomy to be genuinely useful while keeping a human in the loop for anything that could go wrong in a way that matters.

Whether that balance holds up will depend less on Anthropic's own description and more on what happens once a wider pool of Pro and Max subscribers get off the waitlist and start pointing Duet at their actual inboxes, calendars, and spreadsheets.