An AI agent that works while you sleep

Most AI agents live in a browser tab: close it, or close your laptop, and the work stops with it. An agent that is actually still going when you wake up needs three things most chat-based tools don't have, and one honest limit on what it should do without you.

"Works while you sleep" is a nice pitch and a vague one. Nearly every agent that runs inside a chat window fails the literal test: it is bound to your open tab, your laptop's sleep timer, your Wi-Fi. The moment you close the lid, the process behind it is gone, and whatever it was in the middle of is gone with it. For an agent to genuinely keep going overnight, it needs a machine that does not depend on yours.

What overnight work actually requires

1. A computer that outlives your laptop

The agent needs to run somewhere that stays on after you log off — not your browser process, a machine of its own. That is the whole premise of a cloud clone: it has a persistent desktop that keeps running, keeps its files, and keeps its logins whether or not you're watching. See what giving an agent its own computer involves. Without this layer, "runs overnight" just means "runs until your session times out."

2. A list of what it owes you, not just what you said last

A chat model with no memory of commitments will happily forget, by morning, that it was supposed to check a portal at 6am or follow up if a reply hadn't come by then. Working unattended means the agent needs its own record of open items — what it owes you, when it's due, whether you want to be told the moment it changes or only once it's done. On MyClawn that record is a real task store, not the chat transcript: each item has a status, a due time, and a notification policy, and it survives a restart the way a sticky note on your monitor does not.

3. A way to find out what happened, honestly

The riskiest part of unattended work isn't the agent doing the wrong thing — it's you trusting a summary of what it did instead of checking. When you wake up, "I sent the email and it went well" is a claim, not a record. What you actually want is the sent folder, the changed file, the desktop as it was left — and on MyClawn every session can be replayed, the same screen the agent saw, so you can verify rather than take its word. See watching an agent work for what that looks like live, or the recorded replay for what it looks like after the fact.

What we don't claim

Today, checking in means opening MyClawn — there is no push notification to your phone yet when a task finishes overnight. If that matters to you, treat "works while you sleep" as "is still working, and shows you exactly what it did, when you next look" rather than "will wake you up." We would rather say that plainly than imply a notification that doesn't exist.

Where unattended is the wrong setting

• Anything that spends money or signs something.Queue the research, the drafting, the gathering — overnight work is exactly where those tasks belong. Leave the commitment itself, the payment, the submission with legal weight, for a moment when you're there to press the button yourself.
• A task with an ambiguous stopping point. An agent working unattended for eight hours on a task with no clear "done" will find a plausible-looking one and stop there, which may not be the one you meant. Give it a due time and a concrete definition of finished, or expect to correct course in the morning.
• The first time you try any given task. Watch the first run live, or at minimum replay it once it's done. Overnight autonomy is something an agent earns on a task after you've seen it do that task correctly, not something to grant on day one.

What it's good for

The unglamorous, slow, low-risk-per-step work: crawling a set of supplier sites for prices, watching a portal for a status change, drafting replies to a backlog for you to approve in the morning, re-running a check that only matters if something changed. None of it needs your attention while it happens — it needs a machine that keeps running, a record of what it's meant to do, and a way for you to check the work rather than trust it. That combination is described in full at Cloud clones.