Before you let an agent work overnight, decide what it waits on you for.

On September 10, OpenAI opened its Agents API as a public beta. It hands developers the same setup that runs Codex. OpenAI hosts the harness, the agent can keep going for hours, older context gets summarized automatically so a long session doesn't lose the thread, and you choose where it runs: an OpenAI sandbox, your own machines, or one of their sandbox partners. It can call tools, connect to MCP servers, and hand parts of a job to subagents. There's no separate fee. You pay for the tokens and tools it uses.

Not long ago, keeping an agent working through an eight-hour stretch was a project of its own. Now it's a line on an invoice. I think that's good news for small businesses. It also means the hard part of the job has moved.

A night at a small supply business

Say you run a twelve-person supply company. At 11 PM an agent picks up what the day left behind: the support inbox, Stripe payouts to match against invoices, a few vendor bills due Friday, and forty quotes nobody followed up on. By 7 AM most of that should simply be done. Order-status questions answered from tracking numbers. Payouts matched. Tickets tagged. Nobody needs to approve any of it, and if you make someone approve it, you've paid for a queue.

A handful of things from that same night are different. A customer wants $412 back, and your rule is that refunds over $250 get a look. A vendor emailed on Tuesday with new bank details, and their payment is ready to go to the new account. A frustrated customer asked for the owner by name, and the drafted reply offers a credit. The quote follow-up is ready for 38 people, and three of them already told your sales rep no last week.

Each of those is one step away from money leaving, a promise being made, or mail going out under your name. The agent can prepare all four. None of them should go out at 2 AM with nobody awake.

Walk through the night

Drag the marker below from 11 PM to 7 AM, or press play. The log fills in as the agent works. The four items that touch money, customers, or outgoing mail go to the desk on the right instead of going out, and you decide what happens to each one. The business, names, and amounts are made up.

What you should see at 7 AM

The morning is where this works or doesn't. You want a short stack, not a transcript of the night. Each item should show you the actual thing: the email as the customer will read it, the refund with the order history next to it, the payment with the bank change pointed out. Then it should let you send it, fix it, or stop it in under a minute. If the stack runs to thirty items every morning, the rules are too tight and people will start approving without reading. If it's empty every morning, check whether the agent is deciding things you would have wanted to see.

Where I'd draw the line to start

The Agents API doesn't decide this for you, and I wouldn't want it to. It depends on your business. Here's the list I'd start a small team with: money going out above a set amount, any change to where money goes, anything that promises a customer something, and any email going to more than a handful of people at once. Everything else runs. Then read the morning stack for two weeks and move the line based on what you actually approved, what you fixed, and what you stopped.

Buying the harness is the easy part now. Before the first overnight run, write down those four or five rules and name the person who reads the stack each morning.