Tech • AI • Robotics • Game
Anthropic Just Dropped DUET: The First AI That Thinks Ahead on Its Own
9.2/10TL;DR
Anthropic has unveiled Duet, a proactive Claude-powered personal work agent that can monitor connected business tools, prepare tasks in the background, and ask for approval before taking sensitive actions.
KEY POINTS
Crowded race for proactive agents
Duet arrived amid a rapid rollout of similar products from major AI companies. Meta launched Muse on September 8, while OpenAI released Dot on September 30, putting three large tech players into the personal-agent market within about six weeks. The shared direction is clear: assistants are shifting from reactive chatbots to systems that initiate work on a user’s behalf.
Launch and availability
Anthropic announced Duet on October 21 as a research preview. It is currently available only through a waitlist for Pro and Max subscribers in the United States, with access offered on iOS, Android, and the web through claude.ai. The product runs on Claude Opus 5.5 and is aimed mainly at professionals and small-business owners.
How Duet works
Duet is designed to review connected tools such as email, calendars, financial platforms and internal business systems, then surface what needs attention. It can draft documents, organize workflows, monitor progress over hours, and continue running after a user closes a laptop or phone because it operates in its own virtual machine with its own browser. Users can interact through one ongoing thread by text or voice.
Business workflow example
In one demonstration, a coffee-company owner received a morning brief with three urgent items already prepared for review: a license application, hiring arrangements, and a contract counteroffer. Duet then offered to build a sales dashboard by pulling data from point-of-sale systems, an online store, a CRM, customer service, payroll and bank accounts. It also flagged that a Chicago store was opening in 26 days and that a health permit had not yet been submitted.
Long-running task management
The agent is built to handle multi-step work across a day rather than only instant replies. In a hiring scenario, it tracked interviews over several hours, summarized interviewer feedback, identified Jen as the top candidate, noted that she had another offer, and then drafted an offer email with the compensation package attached. The final message was held for human confirmation before sending.
Broad tool access
Anthropic says Duet can connect to more than 6,000 tools and plugins through the Claude marketplace. Examples include checking QuickBooks or Stripe for unpaid invoices, running payroll in Gusto, posting approved job listings to LinkedIn, sending agreements through DocuSign, drafting sales follow-ups from HubSpot, creating campaigns in Canva, and managing a Shopify store.
Payments and purchasing
The system can also make business purchases with safeguards. It uses Link, Stripe’s digital wallet, and can check out through Shop Pay. Users must confirm the item, and payment approval is handled separately; purchases use a virtual card so a real card number is not shared with the merchant.
Early user anecdotes
Internal and early-access users highlighted the appeal of proactive monitoring. A real-estate broker said Duet caught a property showing scheduled on the wrong day without being asked. A restaurant owner in South Florida with about 110 employees used it across accounting, customer support and marketing, saying it reduced the need to juggle “100 tabs.” A brand strategist said the system noticed a client call scheduled right before a flight and suggested moving it.
Safety limits and approval rules
Anthropic says Duet can read, draft and prepare actions autonomously, but it must seek approval before sending messages, making payments, signing documents or deleting anything. Some actions remain blocked even with consent, including trading securities or moving money between accounts. The company says the system does not directly access stored passwords or sign-in tokens and includes protections against prompt injection, data leakage and hidden malicious instructions.
Limits and trade-offs
Duet does not monitor accounts continuously; it checks connected apps every few hours to avoid exhausting plan limits. Background checks and larger jobs consume usage capacity differently from a standard chatbot session, which may matter for heavy Claude users. Anthropic also cautions that the product is early, may misjudge urgency, is stronger on work tasks than personal errands, and does not yet let users watch its browser activity live.
CONCLUSION
Duet signals a new phase in AI competition, where the key battleground is not just model quality but trusted autonomy across email, finance, scheduling and operations. Its success will likely depend on whether users accept an agent that acts first while still keeping humans firmly in control of consequential decisions.
Ask a question