What Is Claude Duet? Inside Anthropic's Always-On AI Assistant

Signup Write New Story What Is Claude Duet? Inside Anthropic's Always-On AI Assistant by Sebastian Buzdugan by Sebastian Buzdugan | @ sebuzdugan ML Engineer | PhD Student in AI

Anthropic just shipped an agent that does not wait for your prompt. It reads your inbox every few hours, drafts the work, and pings you only when it needs a yes.

Every Claude product so far has been reactive. You open a chat or a terminal, you type, it answers.

Claude Duet flips that. Anthropic opened it as a research preview on October 21, 2026, and the pitch is one line: an agent that "comes to you with what needs your attention."

It hit the top of Hacker News the same evening with 470 points and over 400 comments. Most of the thread was not about the model. It was about trust, limits and who owns the agent.

That is the interesting part for engineers, so that is what this piece covers.

Duet is a standalone app on iOS, Android and the web. You talk to it in one long thread, by text or voice.

There is no project setup and no prompt engineering. It carries over the memory, skills and connectors you already have on your Claude account.

Under the hood it runs on Claude Opus 5.5, and it is included in Pro and Max at no extra cost. The catch: it draws from the same usage limits as the rest of Claude.

Access is a waitlist, US Pro and Max subscribers first. Team and Enterprise are not included yet.

The key design choice is where Duet runs.

It does not run on your laptop or phone. Anthropic gives each user a virtual machine in the cloud with its own browser, so the agent keeps working when your devices are off.

From there it reaches your apps through connectors. Anthropic says it can use more than 6,000 connectors and plugins, naming Gmail, Slack, Stripe, QuickBooks, Gusto, HubSpot, Docusign, Canva and Shopify.

One Hacker News user asked Duet about itself, and it described each session as a cloud Linux container with a shell, code tools and a browser. It refused to list its files or its instructions.

That last detail matters. You get the output of the machine, not a view into it.

Duet is built around a polling loop, not a live stream.

It checks your connected accounts every few hours, not every minute. Anthropic is open about why: checking constantly would burn through your plan limits much faster.

Each morning it sends a brief: a short summary of what matters across your apps. During the day you hand it tasks, it works in the background, and it messages you when something is done or needs a decision.

Results come back as real artifacts: a doc, a sheet, a deck.

If you have ever built this yourself, the shape will look familiar. It is a scheduled agent with memory and a notification channel.

The part I would study closely is the permission design, because it is the hard problem in any agent that touches real accounts.

Purchases go through Stripe's Link wallet with a virtual card, so the agent never handles your real card number. It also never directly sees your passwords or sign-in tokens.

On top of that sit two classifiers. One auto-approves safe actions, the same kind used in Claude Code's auto mode. A separate one watches for data leaks and rejects credit card numbers and government IDs outright.

Anthropic also says the app was red-teamed, pen-tested by an outside firm, and is in scope for its public bug bounty.

This is the pattern worth copying for your own agents: read freely, gate every irreversible write, and hard-block the actions you never want automated.

Anthropic lists its own limits, which I appreciate.

It can misread a thread or misjudge what is urgent. Because it polls, it can miss something that lands between check-ins. It cannot show you its browser yet, and it lives in a separate app from the main Claude chat.

The Hacker News thread added a longer list:

One commenter who tried it said the suggestions after setup were mediocre, and that it dumped raw tool-use cards into the chat without saying why.

None of this is fatal for a research preview. It does tell you what to watch in the next release.

Plenty of engineers already run a homemade version of this: Claude Code on a schedule, a few MCP servers, and a message to their phone.

A minimal DIY version is just a cron job that runs Claude Code in headless mode with a fixed prompt:

# crontab -e: morning brief at 07:30 on weekdays 30 7 * * 1-5 cd ~/agents/brief && claude -p "$(cat brief-prompt.md)" --output-format text > ~/briefs/$(date +\%F).md That gets you a scheduled run, full visibility into the prompt and files, and any model or connector you want. What it does not get you is the approval gates, the credential isolation, the cloud VM and the mobile thread.

If you want an assistant that runs your inbox and calendar with zero setup, Duet is built for you. If you need custom check intervals, niche connectors, inspectable rules or a different model, the DIY route still wins.

Duet is less about a new model and more about a new default: the agent starts the conversation.

The engineering bets are clear. A cloud VM so it never sleeps, polling instead of streaming to protect your limits, and a hard approval wall in front of anything you cannot undo.

The open questions are just as clear. How much of a Pro plan does a full day of Duet use, and will you ever get to see what the machine is doing?

Until those are answered, treat it like a new hire on probation: let it read and draft everything, and keep your hand on the send button.