# Claude Duet joins the always-on agents, with a list of things it won't do even if you approve URL: https://www.thedeepfeed.ai/posts/2026-10-21-claude-duet-draws-a-hard-line/ Category: Agents Published: 2026-10-21 Author: the-deep-feed Tags: anthropic, claude-duet, agents, always-on-agents, approvals, consumer-ai Kind: news > Anthropic's Claude Duet, opened as a US research preview on October 21, copies the hosted-computer design of Muse and Dots but forbids some actions even with the user's approval and checks accounts every few hours, not continuously, to stay inside existing plan limits. ## TL;DR - **Claude Duet** opened on **October 21** as a waitlisted research preview for US **Pro and Max** users on iOS, Android and claude.ai, built around **Claude Opus 5.5** and running on a hosted virtual machine with its own browser. - It asks before it sends, pays, signs or deletes, and some actions, such as **trading securities** or **moving money between accounts**, are off-limits **even with approval**. - Duet checks connected accounts **every few hours** rather than continuously, because, Anthropic says, constant checking would burn through plan limits. Its background checks count against the same limits as the rest of Claude. - The launch-day Hacker News thread drew **568 points** and **537 comments**; a follow-up eight days later calling it the first useful assistant harness drew **113 points**. Anthropic's Claude Duet is the fourth always-on personal agent from a major lab since August, and it uses the same design as the others, with one difference: some actions are refused even when the user approves them. It also checks in every few hours instead of continuously, which keeps the product inside existing subscription limits. Anthropic [opened Duet](https://claude.com/blog/claude-duet) on October 21 as "a research preview of Claude Duet, a new app where Claude works as a proactive agent." The [release notes](https://support.claude.com/en/articles/12138966-release-notes) list it the same day. It runs as a standalone app on iOS and Android and on claude.ai, and access starts "with Pro and Max users in the US" through a waitlist. Anthropic says it "designed it around Claude Opus 5.5." # The same stack as Muse and Dots This publication [described the pattern on October 13](/posts/2026-10-13-always-on-agents/): between August 11 and September 29, SpaceXAI's Grok Bot, Meta's Muse and OpenAI's Dots each shipped a vendor-hosted computer per user, agents that keep working after the app closes, and a reviewer model that screens actions. [MacRumors](https://www.macrumors.com/2026/10/21/anthropic-claude-duet/) called Duet Anthropic's "answer to Meta's Muse agent and OpenAI's Dots agent," and the parts list matches. "Claude Duet has a virtual machine with its own browser, so it doesn't require your laptop or phone to be on," the launch post says. Credential handling "follows the design we use for Claude Managed Agents, so it never directly sees your passwords or sign-in tokens." It reaches outside services through "over 6,000 connectors and plugins in the Claude Marketplace," and the launch post names ten, among them Gmail, Slack, Stripe and Shopify. Purchases go through Stripe's Link, and "It then pays with a virtual card." The routine is set by the agent, not the user. "Each morning, it sends you a short brief on the most important items across your connected apps. Throughout the day, it keeps tabs on those apps and messages you only when something needs your attention." Like its rivals, it relies on a second model for screening: Anthropic writes that "Actions are approved by our auto mode classifier." # Where Duet draws a harder line The difference is in what the person cannot hand over. Anthropic says that "before it sends a message, makes a payment, signs a document, or deletes anything, it shows you exactly what it's about to do and waits for your approval." Then it goes further: "A few things are off-limits even with your approval, like trading securities or moving money between accounts." That is a change in kind. Dots, as of September, kept money transfers on a short list of steps it hands back to the person. Duet removes that category from the product altogether. [ITdaily](https://itdaily.com/news/software/claude-duet/) put it plainly: "Certain actions, such as transferring money, are prohibited even with approval." A user cannot delegate the riskiest financial actions to Duet at all, and Anthropic does not have to adjudicate a disputed transfer the user signed off on while half paying attention. Duet also has a dedicated classifier and safety system prompt "designed to reject requests to share the strictest tier of information, including credit card numbers and government identification." In practice, payment details pass through Stripe's virtual card instead of through the model's context. # Always-on, every few hours The other choice is cadence. "Claude Duet checks your connected accounts every few hours, because checking nonstop would use up your plan's limits much faster." Duet is "included in Pro and Max plans and uses the same usage limits as the rest of Claude," and "Its background checks count toward those limits." This is a pricing decision presented as a product decision. In the October 13 survey, Dots shipped inside ChatGPT Pro tiers priced at $100, $200 and $500 a month, and Muse sold a weekly token allowance. Anthropic did not create a new tier. It rationed attention instead. An agent that looks every few hours is cheaper to run and easier to fit under an existing plan, and it is also slower to notice the urgent email that a proactive assistant is supposed to catch. Anthropic lists the limits itself: Duet "still might misread a thread or miscalibrate what's urgent," "personal errands, such as booking tickets, aren't quite as strong, and it can't yet show you its browser." ITdaily noted that Anthropic "has not commented on a launch in other markets or general availability." # The reception The launch-day [Hacker News thread](https://news.ycombinator.com/item?id=50186037) reached 568 points and 537 comments. The most common reservation was trust: > These agent assistants are really appealing to the optimist within me, but I have a hard time making the trust leap. > > — charlesabarnes, Hacker News, Oct 21, 2026 Others questioned the packaging, with one commenter arguing it should "just be native to Claude, not a separate app/UX." Eight days later, a [second thread](https://news.ycombinator.com/item?id=50292727) titled "Claude Duet is the first personal-assistant-type harness that feels useful" drew 113 points and 112 comments. The design question Duet raises is whether a hard refusal list will become the norm or stay a differentiator. The reviewer-model pattern lets vendors sell more autonomy and leave the consequences with the user. A deny list that no approval can override is the first sign of a vendor deciding that some consequences are not worth selling at all. ## Sources - [Anthropic — Introducing Claude Duet (Oct 21, 2026)](https://claude.com/blog/claude-duet) - [Claude Help Center — Release notes, October 21, 2026](https://support.claude.com/en/articles/12138966-release-notes) - [MacRumors — Anthropic Launches Claude Duet, a Proactive AI Agent (Oct 21, 2026)](https://www.macrumors.com/2026/10/21/anthropic-claude-duet/) - [ITdaily — Anthropic launches Claude Duet: proactive assistant arrives late to the party (Oct 22, 2026)](https://itdaily.com/news/software/claude-duet/) - [Hacker News — Claude Duet discussion (Oct 21, 2026)](https://news.ycombinator.com/item?id=50186037) - [Hacker News — Claude Duet is the first personal-assistant-type harness that feels useful (Oct 29, 2026)](https://news.ycombinator.com/item?id=50292727) --- Canonical: https://www.thedeepfeed.ai/posts/2026-10-21-claude-duet-draws-a-hard-line/ Site: https://www.thedeepfeed.ai Full corpus: https://www.thedeepfeed.ai/llms-full.txt