Claude Duet: What Anthropic's Proactive Agent Actually Does

Anthropic opened a research preview of Claude Duet on October 21, 2026, and it is a different shape of product from the chat window most people know. Instead of waiting for a prompt, Duet watches the apps you connect, decides what needs you, and messages you first. That inversion is the whole pitch, and it is also where the risk sits.

The short version: Duet is a proactive agent built on Claude Opus 5.5, it runs on a virtual machine with its own browser, and it stops to ask before it sends, pays, signs or deletes anything. It is a waitlist preview, it shares your plan’s usage limits, and at launch it was offered to Pro, Max and Ultra subscribers in the United States.

What Claude Duet actually is

Duet ships as a standalone app on iOS and Android, and it also runs on claude.ai. Anthropic describes it as being built for “professionals and small business owners with more on their plates than they can track”, which is a polite way of saying it is aimed at people who lose track of email, invoices and paperwork rather than at hobbyists.

It is not a new model. Anthropic built it around Claude Opus 5.5, the same model that anchors the rest of its recent lineup, and it carries over your memory, skills and connected apps from your existing Claude account.

The morning brief is the whole idea

Every morning Duet sends a short summary of the most important items across your connected apps. Through the day it keeps checking those apps, but not constantly: Anthropic says it checks “every few hours, because checking nonstop would use up your plan’s limits much faster”. You get a message only when something crosses the line Duet has drawn for urgency.

That cadence is worth understanding before you sign up. This is not a live assistant listening at all times. It is a scheduled reader with judgment, and the judgment is the part Anthropic itself says is still being tuned.

When a task crosses from reading into doing, Duet opens a virtual machine with its own browser, works out which apps to use, and runs the steps in the background. Your laptop or phone does not need to be on. You get a notification when the task finishes or when it needs a decision.

What it can reach into

Anthropic puts the number at more than 6,000 connectors and plugins in the Claude Marketplace. The examples it gives are concrete and mostly small business flavoured:

Finding missed Gmail and Slack messages and drafting replies.
Checking Intuit QuickBooks or Stripe for unpaid invoices and drafting reminders.
Preparing and running payroll in Gusto with hours, bonuses and reimbursements filled in.
Drafting a job posting, then posting it to LinkedIn once you approve and edit it.
Preparing agreements and sending them for signature through Docusign.
Drafting follow-ups from contact and deal history in HubSpot.
Turning a product launch into a social campaign in Canva.
Setting up and managing a Shopify store, with a richer visual editor promised.

Purchases go through Stripe’s Link wallet or Shopify’s Shop Pay. Duet confirms the item with you, Link asks you to approve the spend, and the payment runs on a virtual card so the real card number is not handed over. Support for logins saved in 1Password is described as coming soon.

The approval rules that matter

Anthropic is unusually specific here, and this is the part worth quoting back to anyone who assumes “agent” means “unsupervised”. Duet reads, drafts and prepares work on its own. Before it sends a message, makes a payment, signs a document or deletes anything, it shows you exactly what it is about to do and waits for approval.

What it will not do, even with your permission

Some actions are off the table entirely. Duet will not trade securities and will not move money between accounts, approval or not. You can also disconnect any app it has access to at any time, which is the practical escape hatch if the integration list gets ahead of your comfort level.

How the security design works

The approach is assembled from pieces Anthropic already runs elsewhere rather than invented for this product. Credential handling follows the design used for Claude Managed Agents, so Duet never directly sees your passwords or sign-in tokens. Actions pass through the same auto mode classifier that guards Claude Code, defences against hidden instructions come from Anthropic’s prompt injection research, and the whole thing was penetration tested by red teams and an external firm under a public bug bounty.

There is one new piece. Duet ships with a dedicated classifier and safety system prompt that screens each message, built to reject sharing of the strictest tier of information, including credit card numbers and government identification. Health and personal finance data are allowed through only when you have explicitly told it to send them. Given that the product’s job is deciding where information should go, that classifier is doing more work than it sounds like.

The limits Anthropic admits to

Anthropic is fairly candid that this is early. Duet “might misread a thread or miscalibrate what’s urgent”, it was built around work so personal errands such as booking tickets are weaker, and it cannot yet show you its browser, which means you cannot watch it work. It also lives in its own app, separate from the main Claude app, and background checks count against your plan’s limits, with larger tasks burning through more of them.

Availability: where Canadian users stand

At launch the preview was opened to Pro, Max and Ultra subscribers in the United States, with a waitlist rather than instant access. Anthropic has not published a Canadian date. If you are already paying for a Claude plan in Canada, the honest position is that you are waiting on an announcement, not on a setting you have missed.

Confirmed versus not confirmed

Confirmed by Anthropic: the October 21, 2026 research preview; standalone iOS and Android apps plus claude.ai; inclusion in Pro, Max and Ultra plans; Opus 5.5 as the underlying model; the approval requirements; the securities and money-transfer prohibitions; the 6,000-plus connector count; the US-only launch.

Not confirmed: any Canadian launch date, pricing outside existing plan tiers, whether the 1Password logins support has shipped, and whether the separate-app arrangement survives past the preview. Treat any claim about Canadian availability as unverified until Anthropic says otherwise.

What to watch

The interesting question is not whether Duet can draft an invoice reminder. It is whether the urgency judgment holds up when the inbox is messy, because a proactive agent that cries wolf gets muted in a week. Anthropic says it will keep tuning that behaviour from tester feedback.

If you want the background on the model underneath, our breakdown of what changed in Claude Opus 5.5 and what it costs covers the pricing side that Duet inherits.