Until now, almost every AI assistant worked the same way: you open a chat, you ask, it answers. Close the tab and the work stops. On October 21, 2026, Anthropic launched a research preview of a product that flips this around: Claude Duet, an agent that takes the initiative.
It sends a brief every morning, watches your apps during the day, handles big jobs while you are busy with something else and only comes back when it needs a decision. Here is our read of what it is, how it works inside and what it says about AI agents working inside a business.
Where this comes from
Duet is in preview, waitlist only, for Pro and Max subscribers with a US account. Details may change.
What Duet does day to day
Anthropic describes the audience as “professionals and small business owners with more on their plates than they can track.” A day with it has three moments.
- In the morning, a brief. It reads email and calendar, flags priorities and starts the first steps, such as drafting replies for you to review.
- During the day, check-ins. Every few hours it looks at your connected apps and only reaches out when something needs attention.
- When you ask, big jobs. You describe the goal, it picks the apps, estimates the effort and returns editable files: a financial model, a deck, a dashboard, a prototype.
Anthropic’s own examples are pure operations: finding unpaid invoices and drafting reminders, preparing payroll for review, building a company brief before a meeting, a daily revenue and cash dashboard with anomaly alerts.
What kind of harness it is
A harness is everything around the model: the loop that calls it, the tools, the memory, where it runs and who checks what it does. Putting together what Anthropic has published, Duet’s harness can be described clearly.
- A single agent. One Claude Opus 5.5 per person, in one continuous thread, with topic channels that share the same memory. It is not a team of role-based agents.
- In the cloud, not on your device. It has its own cloud computer and, separate from it, a machine just for the browser. That is why it keeps working with your phone off.
- Woken by a clock and by messages. It wakes for the morning brief, for check-ins and when you write. It does not react in real time: anything arriving between check-ins waits.
- Effort that adjusts. There is no model choice; it decides how much effort each task deserves.
- Three kinds of hands. Claude account connectors (Gmail, Calendar, Slack, Stripe, QuickBooks, HubSpot and more), a browser for sites without a connector and the computer for producing files.
One important detail: memory, skills and connectors come from the Claude account. Anyone already using Claude starts with an agent that knows who they are.
The agent loop, step by step
Take a job like “find unpaid invoices and chase the customers.” Based on Anthropic’s description, the path looks like this:
- Trigger: the afternoon check-in or your request.
- Context: loads the thread, memory and skills.
- Plan: picks the apps and the effort.
- Action: reads Stripe, matches it against email and drafts the reminders.
- Action check: a separate classifier checks whether each step matches what you want.
- Approval: sending messages, paying, signing and deleting always stop at a card with the exact content. Until you approve, it gets on with the rest.
- Outbound check: another classifier blocks card numbers, bank accounts and ID numbers.
- Delivery: posts the result, citing where every number came from, and goes back to sleep.
Where the safety lives: outside the model
Duet’s most interesting engineering decision is not trusting the model to keep secrets. Instead of teaching the agent to protect the password, Anthropic put the password out of its reach.
- Website passwords are typed into an in-app form, encrypted with a one-time key that exists only on the browser machine and pass through the agent sealed.
- Connector tokens live in a vault. The agent holds a placeholder, and a proxy adds the real token only at call time. It is the same design as Managed Agents, the agent infrastructure Anthropic sells to companies.
- Purchases go out on a single-use card (Stripe’s Link or Shop Pay), valid only for that merchant and that amount.
- The browser only takes fixed commands (open, click, type); cookies and site code stay on its machine.
The approval card and the classifiers sit outside the model doing the work. An email with hidden instructions might confuse the agent, but it cannot approve anything on your behalf.
Some actions are blocked even with approval: trading securities, moving money between accounts and accessing someone else’s account. And paying or creating an account can never be set to “don’t ask again.”
The business model
Duet has no price of its own. During the preview it is included in Pro ($20 a month) and Max (from $100), but it draws on the same usage limits as the rest of Claude, including background work you never asked for.
In our view, that makes the proactive agent a reason to upgrade: real users burn through limits several times a day. It also ties users to the ecosystem, because memory and connectors live in the account.
What it doesn’t do yet
Anthropic is candid about the preview’s limits, and they say a lot about what an agent still needs to work inside a company.
- You cannot watch or take over its browser.
- It checks at intervals, so something urgent may wait.
- No local files, no phone calls, no text messages.
- Microsoft 365 is not available yet and Slack is limited.
- No team or enterprise plans, and it only opens in the US.
- It can misread a thread or misjudge urgency.
Early reviews say the same: the launch is beautifully finished, but it still stalls on real tasks such as inbox triage and sales decks.
What this means if you want agents in your operations
Duet is the world’s largest AI company saying, in a shipped product, that the next step is an agent that works without waiting to be asked and checks in before anything consequential. Its examples are operational: collections, payroll, scheduling, dashboards.
At the same time, its limits show where the real work is: connecting the agent to the company’s internal systems, its ERP, its messaging channels and Microsoft 365, designing who approves what, handling GDPR and similar rules, and doing it for the whole team, not just one person.
- The AI Analyst maps your routines and picks where to start.
- The Forward Deployed Engineer builds the agent inside your operations, with a credential vault, approvals in the right places and a record of what was done.
- Training gets your team ready to delegate and approve safely.
Collections, payroll and scheduling are just examples. If a task repeats, it can be automated.
The agent that shows up before you ask is already here. What separates a good agent from one that stalls is the company context around it.
Enjoyed this article?
Get the next one before it hits the blog.
A short email per issue with what actually matters about AI agents at work. No spam.