Your laptop is closed. Your phone is charging in the kitchen. Somewhere in a data center, your assistant just found the only 40-minute gap three calendars share, booked the meeting, and drafted a brief for your 9 AM. And it will wait for your yes before it sends a single email.
That is the pitch behind Claude Duet, the personal agent Anthropic announced this week. It is a standalone app built around Opus 5.5, and the chat interface is the least interesting part. Duet runs on its own cloud computer and keeps working whether or not you are online. Close everything, go to sleep, wake up to a morning brief that was assembled at 5 AM from your calendar, your inbox, and whatever you asked it to track.
What Duet actually does
Strip away the launch-page language and Duet is a proactive assistant with four working parts:
- A cloud computer it owns. Duet does not run on your laptop. It runs on Anthropic’s infrastructure, which is what makes the “works while you sleep” part literal rather than marketing.
- Persistent memory and skills. It inherits memory, skills, and connectors from the Claude app. If you already use Claude, Duet starts out knowing you.
- Payments. Duet can buy things using a one-time virtual card through Stripe’s Link wallet. Not hypothetical, either: it shipped with the preview.
- Sign-off gates. Any deletion, payment, legal signature, or outgoing message stops and waits for your explicit approval first.
The last one sounds like a limitation. It is the feature. More on that below.
Usage counts against your normal Claude limits, so Duet is not a separate subscription. The research preview is rolling out in waves to waitlisted US Max and Pro subscribers, so if you are reading this from Jakarta or Berlin, you are waiting.
The 24/7 part changes what an agent is
Every “AI assistant” you have used until now has been reactive. You type, it responds, the session ends when you close the tab. The assistant has no opinion about your Tuesday.
Duet inverts that. It watches the calendar, notices the conflict, and proposes the fix before you ask. Morning briefs, meeting booking, tracking threads you care about. The grunt work of a workday, handled by something that never sleeps and never gets bored.
This is a bigger deal than it sounds, because the value of an assistant was always capped by how much attention you could spare to direct it. An assistant you must operate is a tool. An assistant that operates itself is staff.
The safety model is the real product
Here is where Duet diverges from most of the personal agent market. Plenty of competitors will happily fire off emails autonomously, because shipping an impressive demo matters more than shipping a careful one.
Anthropic drew the line exactly where the damage lives. Duet can research, draft, schedule, and prepare — but the irreversible actions all stop at a gate. Deleting your files, spending your money, signing your name, speaking in your voice to another human: those require you to click yes.
I find this the most reassuring design decision in the current agent wave, and also the most practical. The failure mode of personal agents is not that they are dumb. It is that they are confidently wrong in ways that send a half-baked email to your boss at 2 AM. A gate on outgoing messages removes the entire category of 2 AM mistakes.
The payment design shows the same thinking. A one-time virtual card means Duet can complete a purchase without ever holding a reusable key to your wallet. That is bank-grade paranoia applied to an assistant app, and I mean that as a compliment.
Why the model underneath matters
The personal agent space is crowded and mostly undifferentiated: same chat box, same connectors, different logo. Duet’s actual moat is that it is the only agent built entirely around Opus 5.5, probably the strongest model shipping right now.
That matters for an agent more than for a chatbot. A chatbot that misreads your intent wastes a message. An agent that misreads your intent books the wrong meeting, emails the wrong client, or buys the wrong thing, at scale, while you sleep. Agent quality is downstream of model quality, and Anthropic built the agent directly on its best model rather than letting third-party wrappers rent it.
There is a wrinkle worth noticing here. Grok Bot already routes some of its harder tasks to Claude. If competitors lean on Claude anyway, Duet simply cuts out the middleman.
How to use a personal agent without getting burned
Whether your waitlist invitation arrives this week or in the next wave, the onboarding decisions matter more than the app. A few rules worth keeping:
Start with read-only work. Morning briefs, tracking, summarizing, research. Let the agent prove it understands your context before it touches anything with consequences.
Keep the gates on. If a future setting offers “fully autonomous mode,” leave it off for a month minimum. Every autonomy horror story starts with someone switching off the checkpoint that was designed to save them.
Watch your limits. Duet counts against your Claude usage, and an agent that runs all night can eat a subscription. Check the usage page before assuming the app is free to run wild.
Give it a real job on day one. Agents earn trust by finishing things. One concrete recurring task, say the morning brief or calendar triage, teaches you more about Duet than a week of open-ended experimentation.
What to watch
Three questions will decide the preview period: how generous the usage limits feel once the agent runs daily, whether the sign-off flow survives contact with real impatience, and when the waitlist opens beyond the US.
The first two determine whether Duet is a real assistant or a very good demo. The third is just a matter of time.
What to do now
Join the waitlist if you are a US Max or Pro subscriber. This is the first personal agent from a major lab whose safety model matches its ambition, and the research preview is where the pricing and limits get figured out. Not in the US, or not a subscriber? Do the prep work anyway: write down the three recurring tasks that eat your week. That list becomes your onboarding brief for whatever agent you end up with, Duet or otherwise.
