Anthropic rolls out Claude Duet, an agent that works while you're away

For the past three years, generative AI has mostly waited around for someone to type a prompt. Anthropic is now betting that you would rather have it work in the background.

Anthropic rolled out Claude Duet, a proactive AI agent app built on Claude Opus 5.5. The app is available as a research preview via a waitlist for US Pro and Max subscribers across iOS, Android, and the web.

A coworker with its own virtual computer

Unlike traditional chatbots, Claude Duet runs on a dedicated virtual machine equipped with its own headless browser. That means you can assign a project, shut your laptop, and let the agent finish the steps on its own.

To keep usage within plan limits, Duet checks in on your apps every few hours rather than running constantly. Every morning, it delivers a brief highlighting the most pressing tasks across your accounts.

Throughout the day, it stays mostly quiet, pinging you only when it needs attention. The agent hooks into over 6,000 integrations via the Claude Marketplace.

Those tools let Duet catch missed threads in Gmail and Slack, draft invoices in QuickBooks, prep payroll in Gusto, or build marketing assets in Canva. Early testers running local businesses reported using it to coordinate store operations without jumping between a dozen open browser tabs.

Guardrails before actions

Giving an AI agent free rein across work accounts requires strict guardrails. Anthropic built in several checks before Duet can touch anything permanent.

While the agent can draft text and compile spreadsheets autonomously, it cannot send an email, sign a document, delete data, or spend cash without explicit user approval. When paying for supplies through Stripe Link or Shop Pay, it relies on virtual cards to hide your actual banking details.

Anthropic also put hard limits in place. Duet is completely barred from trading securities or moving funds between accounts.

A dedicated classifier monitors each message to block leaks of sensitive data like government IDs and credit card numbers. Anthropic also admits the current preview is far from perfect: it is not built for personal errands like booking flights, and it can still misjudge what counts as urgent.

Why it matters

The AI race is rapidly shifting from reactive chatbots that answer questions to autonomous agents that handle workflows. As TechCrunch pointed out, Duet lands right in the middle of a crowded arena, taking on OpenAI's Dots, Meta's Muse, and Grok Bot.

For small business owners and busy operators, Duet could offer genuine leverage. Instead of acting as another chat window you have to babysit, it turns fragmented SaaS apps into a single monitored dashboard, which makes it worth testing if you are already on a paid tier.

We think Anthropic was smart to focus squarely on business workflows rather than personal errands. Business software has clearer pipelines and customers who are already used to paying subscriptions.

That said, leaving an agent in charge of business communications comes with real operational risk. Background polling and multi-step tasks eat through usage limits quickly, and misreading the urgency of a client thread could have real consequences. Keeping a human in the loop for every major click is safe, but it also shows that autonomous software still requires constant supervision.