Anthropic has joined the race to build personal agents that take the first step. On October 21, 2026 the company opened a research preview of Claude Duet, a separate app in which Claude watches a user's connected work tools, decides what deserves attention and carries out tasks in the background. Access starts through a waitlist for Pro and Max subscribers in the United States, and Anthropic says it built the product for professionals and small business owners who have more to track than time allows.
Duet runs as a standalone app on iOS and Android and inside claude.ai. Anthropic says it designed the agent around Claude Opus 5.5 and tuned it to judge what needs the user now, what can wait and when to check in. Each morning it sends a short brief drawn from connected apps, then messages the user during the day only when something requires a response. Tasks can range from booking meetings to building dashboards, preparing financial or sales forecasts, prototyping and making web apps.
The agent inherits a user's memory, skills and app connections from their Claude account, and conversations happen in one continuing thread by text or voice. Work does not depend on the user's own devices: Duet has a virtual machine with its own browser, so it can keep going after a laptop is closed. Anthropic says it checks connected accounts every few hours rather than continuously, because constant polling would drain plan limits faster. Those background checks count against the same usage allowance as the rest of Claude, and larger jobs such as a financial model consume more.
Anthropic is leaning on integrations to make the agent useful. Duet can draw on more than 6,000 connectors and plugins in the Claude Marketplace, and the company's examples include chasing unpaid invoices in QuickBooks or Stripe, running payroll in Gusto, drafting job posts for LinkedIn, routing agreements through Docusign, writing follow-ups from HubSpot records, building Canva campaigns and managing a Shopify store. It can also buy things through Stripe's Link wallet, which pays with a virtual card, and through Shop Pay. Support for signing in with 1Password credentials is promised later.
Control is the central design question. Duet reads and drafts on its own, but before it sends a message, makes a payment, signs a document or deletes anything, it shows the planned action and waits for approval. Trading securities and moving money between accounts stay blocked even with consent. Anthropic says the agent never directly sees passwords or sign-in tokens, that an auto mode classifier vets its actions, and that a new classifier with a dedicated safety prompt refuses to share card numbers or government IDs while releasing health or financial details only on explicit instruction. Red teams and an outside firm tested the infrastructure, and the product falls under Anthropic's bug bounty.
The launch puts Anthropic alongside Meta's Muse and OpenAI's Dots, the proactive agents The Verge and VentureBeat name as Duet's closest rivals. VentureBeat notes the announcement offers no Team or Enterprise access, no Duet API and no enterprise schedule, while Engadget reports Anthropic will share more about Duet's safety protections in the coming days.
Anthropic is candid that this is an early build. It says Duet may misread a thread or misjudge urgency, handles personal errands such as ticket booking less well than work tasks, and cannot yet show users its browser. The company has not published benchmarks or measured productivity results for the agent, so its usefulness will rest on how testers judge its interruptions and the quality of the work it hands back.