An overnight task: handing an AI employee work that runs while you sleep, and waking up to a result rather than a surprise

The cheapest working hours of an AI employee are the night: you are asleep, the accounts are open, nobody is waiting for a reply in the chat. Owners ask search engines whether an AI can run tasks inside their Ozon seller account without them, and whether there is an AI assistant for 1C you can hand a task to overnight. It can, and below we show how. But an overnight task differs from a chat request not in length but in one fact: you are not there. Nobody clarifies, nobody says "stop", nobody notices the agent going the wrong way. Below: how to phrase work that runs without you, where its results arrive, where the agent must stop, and what you will see in the morning.

The trick that went round the timeline

Jarred Sumner, the author of the Bun runtime, posted on X how he leaves an agent working overnight with no special modes or loops. The whole assignment fits in one sentence: "I'm going to sleep, and when I wake up I would love to see X result. Post frequent updates to the channel as you make progress." The post collected more than eight hundred bookmarks, and that is the telling number: people bookmark not what they admire but what they intend to repeat.

Another author the same week priced the night shift. Five hundred videos of release takes came out in a week, a hundred and fifty hours. He had an agent pull every transcript, a little over a cent per video, and write the digest instead. The bill came to six dollars. Minutes watched: zero.

Both posts are about programmers, but Sumner's sentence translates into a business owner's language without loss. "I'm going to sleep, in the morning I want a reconciliation of MoySklad stock against the Ozon account across all warehouses, discrepancies as a table." "In the morning I want the list of invoices overdue by more than five days and a draft email to each debtor." "In the morning I want replies to every review of the week as drafts, send nothing." The only difference: overnight a programmer risks code that can be rolled back, while a business owner risks an email to a customer that cannot.

How an overnight task differs from a chat request and from an AI agent on a schedule

We have already covered automating routine tasks that repeat every week. An overnight task is built differently: it is the third kind of assignment, between the chat and the schedule.

In the chat you supply half of the assignment on the fly: "no, only the main account", "without consignment stock", "wrong period". At night that channel is closed, and everything you left unsaid the agent fills in itself. It fills it in honestly and in its own way, and in the morning you get the wrong report. So an overnight task needs the same text as a scheduled one: the source by name, the period and the breakdown, steps with thresholds, the shape of the result, what to do when there is no data or the service is silent, where to deliver. One difference: an overnight task has no second Monday on which to fix the wording. The first run is the only one.

Hence a practical rule: set up the first overnight task in the evening, but start it while you are still there. Five minutes watching where the agent goes for data and what it treats as the result cost less than a morning with a useless table. If the start looks right, leave; from there it runs on its own.

What you call the result

The key word in Sumner's post is not "sleep" but "see X result". He names the thing he wants to receive, not the process the agent should go through. In the morning he opens not an activity log but a finished result and sees at once whether it matches.

For a business it looks like this. Bad assignment: "sort out the stock". Good: "a table, one row per product, where the stock in MoySklad and in the Ozon warehouses differ by more than two units, with your guess at the cause next to it". Bad: "look at receivables". Good: "a list of counterparties overdue by more than five days, amount and date of the last email, plus a draft reminder to each in a separate file".

Three signs of a well-named result:

• It can be checked in a minute. A table with a threshold is checked at a glance; an "analysis" takes an hour.
• It has a shape. A table, a file, a message of up to ten lines. A shape you did not name, the agent will choose for you.
• The empty result is named too. "If there are no discrepancies, say so in one line" separates silence from failure.

The last point matters most at night. In the morning you cannot tell "the agent found nothing" from "the agent could not log in" if both look like the absence of a message.

Where the result arrives, and what arrives on the way

Sumner asks the agent to post updates to a work channel. With us the result of a task arrives wherever suits you: the chat, email, Telegram, MAX or Pachca. An AI agent in Telegram is the most convenient for a business owner precisely here: the morning report sits next to the correspondence with suppliers. But during the night what reaches the messenger is not a chronicle, only what needs you. In the morning the result arrives. At night a question does.

That is a deliberate choice. Updates every ten minutes turn a night shift into night duty: either you wake up for each one, or you mute notifications and miss the single one that was a question. The full chronicle lives in the task's log, and in the morning you can read it alongside the result if the result raises doubts.

Where the agent must stop

Here the night shift parts ways with what a programmer does. Code can be rolled back. An email to a debtor, a written-off delivery, a changed price in the seller account, a payment on an invoice cannot. Everything the employee does in 1C, MoySklad or on a marketplace is split with us into three classes of action: reading, reversible writing and irreversible writing. At night the split works like this.

The card arrives in the messenger. If you are asleep, the task waits. Not forever: the wait has a limit, after which the run stops with an honest line, "did not receive your confirmation", and everything done before the card stays in the report. In the morning you see where the agent reached the boundary and decide for yourself: confirm and let it finish, or rewrite the task so the boundary is not inside it.

Hence the second rule of an overnight task: move everything irreversible into drafts. "Prepare the emails, do not send." "Collect the list of items to reprice, do not change prices." "Create the supplier orders as drafts." Then overnight the agent does all the heavy part of the work, and in the morning you confirm ten drafts in five minutes instead of finding the one card on which everything stalled at one in the morning.

What else can happen overnight

At night it is not only you who is absent. Marketplace accounts go down for maintenance, sessions expire, 1C on the server restarts. An agent that works unsupervised must be able to do three things we have written about separately: keep its place in the task, never perform the same write twice, and stop honestly when the result cannot be proven.

For an overnight task one more time limit follows. A single run with us lives for up to an hour, and if the agent does nothing observable for ten minutes the run is treated as stuck and closed. That is protection from an endless night in which the agent goes in circles and burns your quota. Work that honestly needs more than an hour is better split into two tasks: the first collects, the second, started later, processes what was collected.

Morning: what to read first

The morning report is read in this order.

1. Did the run reach the end. If not, the first line of the report says why: no confirmation received, the account did not respond, time ran out.
2. The result in the named shape. A table, a file, a short message. If the shape is not what you asked for, fix the assignment before the next night.
3. What is left on cards. Drafts and stopped actions. These are your five minutes of decisions.
4. The log, only if the numbers surprised you. The chronicle of every step lives in the task; normally it does not need reading.

If this order repeats for several nights in a row with the same assignment, the task is asking for a schedule, and the night shift turns into automated reporting. A night shift you start for the third evening running is no longer a one-off.

A task text you can copy

An evening assignment for a stock reconciliation, assembled by the rules above.

There is not a single action in this assignment on which the agent can stall at night, and the result is checked in a minute in the morning. Everything that needs a decision is moved into your five-minute morning.

Sources

• Jarred Sumner, post on X about leaving an agent to work overnight: https://x.com/jarredsumner/status/2097331347853643839
• Post on X about a digest of five hundred videos for six dollars: https://x.com/Jasperli0122/status/2097780078247563324
• Our articles: automating routine work, what the owner sees when the AI employee works without them, an AI agent failed halfway through a task