The overnight agent supervisor: a new category for AI coding

An overnight agent supervisor is the layer that lets a coding agent work while nobody is watching and still be worth reviewing in the morning. It holds the environment, keeps the run alive across failures, stops it at a budget you set, and produces evidence you can check without re-running anything. It is a category because the alternative — an agent nobody supervised — has a predictable failure mode and no answer to it.

Why does this need a name?

Because the tools we already have names for solve a different problem. An in-editor assistant makes you faster while you are at the keyboard; that is a real and valuable thing, and it is finished when you close the laptop. A hosted autonomous engineer takes the whole job somewhere else, along with your code and your model choice. Neither of those is "the same agent I use, running while I sleep, on my machine, against my provider."

That third thing turns out to have its own requirements — persistence, recovery, memory, budgets, and verification — which is the usual sign that something is a category rather than a feature.

What has to be true before you can leave an agent alone?

Four things, and the list is short enough to check honestly against any tool you are considering.

• It survives your terminal. The run belongs to a background process, not a window.
• A failure is an event, not an ending. A red build starts a repair with a retry budget.
• It stops on its own terms. A turn, token and wall-clock budget you set before the run.
• Something other than the model decides it is done. Gates are your commands, and they pass or they do not.

Miss the first and you get four hours of nothing. Miss the second and you get a plan instead of a change. Miss the third and you get a bill. Miss the fourth and you get a confident summary of work that was never verified — which is the expensive one, because it is the only failure that looks like success.

The night shift is not a longer prompt. It is a different job, and it needs someone whose job is to still be there at 4am.

What does a supervisor hand back in the morning?

A branch, and the evidence that produced it. Not a chat transcript to read, and not a claim to take at face value — the commands that ran, what they printed, which of the changed lines actually executed under test, and a signature over all of it that fails if anything was touched afterwards.

That is the review we want a developer to do: read a diff, glance at evidence, decide. Not reconstruct what happened overnight from a summary written by the thing that did it.

Is this just autonomy with better marketing?

No — autonomy is one of the four requirements, and on its own it is the dangerous one. An agent that continues without a budget is unattended, not supervised, and an agent that continues without a gate is generating unverified work faster. The word "supervisor" is doing real work here: supervision is the combination of persistence *and* a standard, and both halves have to be present before leaving the room is a reasonable thing to do.

How do I run one tonight?

One command, your own model, nothing uploaded to run:

curl -fsSL https://relayevals.com/harness/install.sh | bash

Then hand over a goal and a budget instead of a prompt, and read the branch in the morning. The quickstart has the first run, what is a coding agent harness covers the machinery underneath, and pricing is short: free forever locally, and paid only when a team wants the work on the record.