Anthropic launches Claude Duet: proactive agent with virtual machine and daily summaries
Anthropic launched a research preview of Claude Duet on Wednesday, October 21, 2026 — a proactive AI agent designed to monitor connected apps, send daily summaries, and carry out work tasks for small businesses and independent professionals. Access is currently via a waitlist and requires a Pro or Max subscription in the US. The launch comes just weeks after Meta's Muse and OpenAI's Dots, completing the picture of a new generation of "proactive" agents that don't wait for commands but keep an eye on your connected apps on their own initiative (CNBC, The Verge).
What Duet actually does
According to Anthropic's announcement, as cited by The Verge, Duet sends a short summary of the most important items across the connected apps every morning. Throughout the day it keeps an eye on those same apps and only speaks up when something needs your attention. In addition, the user can give the agent tasks of varying size — from simple administrative chores like scheduling meetings, to more complex work such as building dashboards with dozens of tools, creating financial or sales forecasts, making prototypes, and building web applications.
For business owners, Duet can, according to Cryptopolitan, go through Stripe for unpaid invoices and draft payment reminders. The agent builds on Claude's existing connector system: CNBC-TV18 reports that it integrates with over 6,000 connectors in the Claude Marketplace, including Gmail and Slack.
Under the hood, Duet runs on its own virtual machine with a browser, so no Claude session needs to stay open while the agent works (Engadget). The model behind it is, according to Anthropic's announcement, Claude Opus 5.5 — but Gizmodo points out that the company only says the agent is "designed around" Opus 5.5, and does not explicitly state that it uses that model exclusively.
The monitoring loop — and why it doesn't run constantly
One detail Engadget reports alone: Duet checks the connected accounts only every few hours. The reason is partly practical — the agent shares usage quotas with all your other Claude tools, including Claude Code. "Duet checks your connected accounts every few hours, because checking nonstop would use up your plan's limits much faster. We're continuing to test its judgment about when to message you, and we'll keep tuning it based on what testers tell us," the company says, according to Engadget.
That means Duet is in practice an agent with built-in throttling: The frequency and topics of the messages are, according to Engadget, to become user-configurable, but initially the reminder rhythm is governed by what conserves the plan's limits. It is a concrete compromise that distinguishes this type of agent from an always-on assistant.
Purchases, payments, and hard limits
Most concrete is the way Duet handles money, described by Cryptopolitan with reference to Anthropic's announcement: The agent shops online with Link, Stripe's digital wallet, and pays via Shopify Shop Pay. The process is that Duet confirms the item, Link asks the user to approve the purchase, and a virtual card pays — so the real card number is never shared. Duet is also never supposed to see the user's password when logging into websites.
For consequential actions — such as sending a payment or deleting something — Duet is initially supposed to contact the user for approval first (The Verge). In addition, according to Anthropic's description, the agent has a dedicated classifier and a safety system prompt "designed to reject requests to share the strictest tier of information," such as credit card numbers.
And some things are forbidden regardless: Even if the user gives permission, Duet is not supposed to trade stocks or move money between accounts.
It is worth emphasizing that these are Anthropic's own security descriptions, not independently verified facts. No third-party security audit has been presented, and Gizmodo notes that Anthropic's security section largely repeats measures already known from Claude and Claude Code.
Access and pricing — the dividing line against Muse and Dots
For now, a Pro or Max subscription and residence in the US are required, and you must join a waitlist. Engadget reports — citing Anthropic — that Pro subscribers (the $20-per-month tier) will get access when the agent reaches general availability. This pricing information has not been confirmed by other outlets, and should be read as a statement from the company.
The price divide is part of the story: OpenAI's Dots are currently limited to OpenAI's Pro tiers, which start at $100 per month. If Anthropic's statement about the $20 tier holds, Duet positions itself markedly broader. There is so far no indication of a free tier (Gizmodo).
Duet is available as a standalone app for iOS and Android as well as through Claude on the web (The Verge).
The race it launches into
The launch takes place in a market that has only just taken shape. Meta's Muse launched last month and topped Apple's App Store. On Meta's third-quarter earnings report, Wednesday evening US time, Mark Zuckerberg told analysts that over 10 million people in the US and Canada "have set up Muse personal agents" in the past month and a half. This is a company statement, not an independently verified figure. OpenAI launched Dots at its developer conference a few weeks after Muse (CNBC).
Against that background, Anthropic's positioning appears deliberately different: Muse is personal and consumer-oriented, Dots are expensive and initially on the top tiers, while Duet explicitly targets small businesses and professionals — with workflows, invoices, and forecasts as its main tasks rather than personal errands.
Caveats and open questions
Anthropic itself is clear that Duet is still early: The company says, according to CNBC, that the agent "might misread a thread or miscalibrate what's urgent." It is also said to be weaker at personal errands, such as booking tickets. These limitations are stated by the company itself — and independent testing of how reliable Duet is in practice does not yet exist.
The open questions are therefore many: How well do the approval gates and the classifier hold up in real attacks or error situations? How does sharing usage quotas with Claude Code work during a busy work week? And does the $20 promise for general availability hold? For now, we only know what Anthropic has promised about what Duet can and cannot do. The waitlist will decide who gets to find out first.