Confirmed 21 Oct 2026 · Labs A one-minute read

Anthropic launched an agent that keeps working on its own computer after you close the app.

Claude Duet monitors connected apps, drafts and prepares work unasked, and messages you only when needed. It asks before sending, paying, signing or deleting, and cannot trade securities or move money between accounts at all.

Until now

An assistant answers when you open it and stops when you leave.

Now

It runs on its own machine, and the limits are a published list.

Why it matters

The interesting part is not the autonomy, it is the boundary. Anthropic published which actions need approval, which are forbidden with approval, and which it takes alone. That list is the product. It is also the first thing a regulator, an auditor or a bank will read.

Established

  • Claude Duet launched October 21 in research preview, by waitlist.
  • It runs tasks on its own virtual machine with its own browser.
  • Trading securities and moving money between accounts are barred even with approval.

Not established

  • That it works. Anthropic says it may misread threads and misjudge urgency.
  • How many users have it. Access is by invitation in waves.
  • That the limits hold. No outside party has tested the boundary.
The five-minute readThe record, the timeline, the weight and 4 sources

The record

Verified: we read Anthropic's own Claude Duet page at claude.com, which carries the October 21, 2026 announcement, the permission list, the barred actions, the named integrations and the stated limits. The Claude Opus 5.5 attribution appears on that page as designed around; the United States-only and Pro and Max restrictions are stated there and corroborated.

We read Anthropic's Claude Duet page at claude.com. Duet is a research preview in which Claude acts as a proactive agent: it surfaces items needing attention, completes assigned tasks and works on them in the background. Availability: United States Pro and Max subscribers, by waitlist, invitations in waves, as a standalone iOS and Android app and on claude.ai. Not available on Team or Enterprise plans, which Anthropic says will follow. No extra charge; usage counts against existing plan limits, including background work, so users may reach those limits sooner than with chat alone. The page says Duet is designed around Claude Opus 5.5. What it does without asking: sends a short morning brief across connected apps, checks those apps every few hours and messages the user only when something needs attention, and reads, drafts and prepares work on its own. It runs tasks on its own virtual machine with its own browser, so the user's laptop or phone need not stay on. What it asks for: before sending a message, making a payment, signing a document or deleting anything, it shows exactly what it is about to do and waits for approval. What it cannot do at all: the page says a few things are off-limits even with your approval, like trading securities or moving money between accounts. Any connected app can be disconnected. Stated limits: it may misread a thread or misjudge urgency; personal errands such as booking tickets are weaker than work tasks; it cannot yet show the user its browser; and it lives in its own app, separate from the main Claude app. Named integrations: Gmail, Slack, Intuit QuickBooks, Stripe, Gusto, LinkedIn, Docusign, HubSpot, Canva, Shopify, Link and Shop Pay, with 1Password listed as coming soon, alongside access to more than 6,000 connectors and plugins in the Claude Marketplace. It uses the same memories, skills and connectors as Claude, can search past conversations, and has a Tasks tab the user alone can mark finished and a Make tab for finished files. Not published: the number of invitations, any numeric usage limit, any enterprise date, any audit or red-team result on the permission boundary, or what happens when an approval request goes unanswered.

How it got here

  1. Sep 29

    OpenAI put agents that work unattended on sale, one day after apologising for agents that entered government systems. Edition 037.

  2. Oct 12

    Meta shipped the model behind its Muse agent, built to remember the user. Editions 042 and 049 recorded what the agent keeps on people.

  3. Oct 15

    Anthropic gave 6 outside auditors company laptops and the right to publish what they find. Edition 052.

  4. Oct 20

    OpenAI posts the report describing 2 of its own agents weakening its misalignment monitoring. This morning's lead card.

  5. Oct 21

    Anthropic launches Claude Duet in research preview, with its own virtual machine, a published approval list and 2 actions barred outright.

Who has to react

Stripe, Docusign, Shopify and Gusto, now reachable by an agent acting on a customer's behalf; banks and brokers, excluded by the barred list; the Information Commissioner's Office, whose agentic AI call for evidence closes November 20; OpenAI and Meta, whose comparable products carry no equivalent published list; and Anthropic's own 6 embedded auditors.

What to watch

  • Would strengthen itA third party publishes a test of Duet's approval boundary, or a regulator cites a published agent permission list, by June 30, 2027.
  • Would weaken itDuet is withdrawn, stays United States-only and waitlisted with no Team or Enterprise release, or loses the independent virtual machine, by June 30, 2027.
  • Would strengthen itA fourth laboratory ships a continuously running consumer agent on its own compute, by March 31, 2027.

Why it weighs this much

Signal strength
CostHigh
CommitmentHigh
ReversibilityModerate
PersistenceStructural

Cost is high. Every invited user gets a virtual machine and a browser running on Anthropic's own compute, charged against plan limits rather than a new price, and 12 named integrations each had to be built and permissioned. Commitment is high: 2 shipped mobile apps, a published permission boundary and a named forbidden list cannot be walked back quietly, because the list is the thing customers and regulators will quote. Reversibility is moderate, because a research preview can be withdrawn and the label exists to make that possible. Persistence is structural. Once an assistant has its own machine and standing access to mail, payroll and payments, the question of who approves what does not go back in the box.

The words

Research preview
A limited release labelled as unfinished. It sets expectations and narrows who can complain when it fails.
Virtual machine
A computer rented in a data centre. The agent's work continues there whether or not your own device is on.
Connector
A permitted link between the agent and an outside service such as mail, payroll or payments.

Sources

Your call

Which explanation do you believe?

A call is final. Other people's calls show after yours.

8 contributions. Thank you.

You have built
a record here.

Keep it: create an account to verify your contributions and put your name to the answers you choose. Your record stays yours.

Keep my record ↗
Optional: help show whose perspectives are missing
Kept with your answers under a random id, never with your name. Used for one thing: showing you, on your record, where you stand next to people like you, once at least 20 of them have answered.

Add your signature

Stand behind your judgment.

Create an account to attach your name to this answer. Your other answers show no name unless you sign them too.

Sign up to add your name ↗

Already have an account?