Scheduling AI Work: Recurring Tasks and Overnight Agents

The most valuable AI work is the work you never see happen. It’s already done by the time you sit down.

Most people use AI in one mode: they ask, it answers. That’s the whole relationship. You open a chat, type a request, read the response, close the tab. The AI is idle until you prod it.

There’s another mode, and it’s where a lot of the leverage is. AI that runs on a clock. It gathers your morning briefing before you wake up, watches a deploy until it finishes, or works through a backlog overnight and hands you a report at nine. You didn’t prompt it. It just ran.

Triggered by time, not by you

This is close cousin to automation, but the trigger is different. Event-triggered automation reacts to something happening - a new email arrives, a row gets added, a form is submitted. (That’s the sibling piece on automating work with AI.)

Scheduled work runs on a clock. Nothing has to happen first. Seven o’clock comes, the job fires.

Both are delegation. But scheduled work delegates something specific: your time. The task that used to need you to show up at a certain hour now happens without you in the room.

Three shapes worth knowing

The recurring digest. Every weekday at 7am, gather something and summarize it. Your inbox overnight, your key metrics, the news in your sector, yesterday’s support tickets. Instead of spending the first 30 minutes of your day assembling context, you start with it already on the page.

The watcher. Check something every few minutes until a condition is met, then tell you. A deploy that you want confirmed before you move on, a number you’re waiting to cross a threshold, an inbox you want flagged the moment a specific sender lands. Claude Code’s /loop, for example, runs a check on an interval until it’s worth interrupting you. You stop babysitting the thing and let it tap you on the shoulder.

The overnight batch. Hand off a queue of work at night and collect the output in the morning. Process a stack of documents, draft replies to a backlog, run an analysis across a dataset. This is the “overnight agents” idea: autonomous scheduled delivery, done while you sleep, ready when you’re back.

How it shows up in real tools

For code, this has always been cron: a scheduler that runs a command on a fixed timetable. Nothing new there.

What’s new is that the thing being scheduled can now be an AI agent. Claude Code’s /schedule creates recurring cloud agents that run on a cron schedule and report back. /loop handles the polling case, where you want a check repeated until something changes rather than at a fixed hour. The mechanics are familiar; what’s different is that the scheduled job can read, reason, draft, and summarize, not just move files around.

Keep your expectations honest. These tools are real and current, but a scheduled agent is still an agent. It does what you’ve told it, with the judgment you’ve given it, and no more.

The guardrail that actually matters

Scheduled AI acts when you are not watching. That changes the stakes.

When you’re in a live chat, you see every output before anything happens with it. A scheduled job has no such checkpoint unless you build one in. So be explicit about the line between what it may do on its own and what it may only prepare for you.

A simple split holds up well:

• Safe to run unattended: gather, summarize, draft, flag, analyze. Nothing here is hard to undo.
• Needs a human first: send, pay, publish, delete, commit. Anything that touches the outside world or can’t be taken back.

Let the scheduled job do the first list freely. For the second, it should stop and produce something for you to approve, not act.

How to start

Pick one recurring task you already do at the same time each day or week. Define exactly what the AI should produce. Then schedule it to land as a draft or a report you review, not as an action it takes.

Run it that way until it earns your trust. Only then consider letting it act on its own, and only for the reversible things.

Every [weekday] at [7am], do the following:

Gather: [what to pull together - inbox, metrics, tickets, news in X]
Summarize: [how - top 5 items, one line each, grouped by Y]
Deliver: [where - email me, post to this channel, write to this doc]
Do NOT: [send replies, change anything, act on it] - prepare it for my review only.

Where this fits

The skill ladder gets you capable. The context stack makes your AI actually useful on your work. Delegate and automate is the last turn: AI that keeps working when you’ve left the room.

Start with one scheduled draft waiting for you tomorrow morning. That’s the whole idea, scaled down to something you can set up today.