Blog
What is an AI agent supervisor?
An AI agent supervisor runs long tasks, recovers from failures, and verifies the result before you review it. How supervision differs from a chat assistant.
2026-09-08 · Sathwik R
An AI agent supervisor is a layer that manages a coding agent the way a lead manages an engineer: it hands over a goal rather than a prompt, keeps the work moving when something breaks, holds it to a standard before calling it done, and reports back with evidence. You delegate to a supervisor. You converse with an assistant.
What is the difference between a supervisor and an assistant?
An assistant answers the question you asked. A supervisor owns an outcome you named. That single difference changes everything downstream — how long the work runs, who notices a failure, what "finished" means, and whether you are reviewing a result or a claim about one.
| Assistant | Supervisor | |
|---|---|---|
| You give it | A prompt | A goal and a budget |
| It stops when | It has answered | Its gates pass, or the budget ends |
| A failed command | Ends the turn | Starts a repair |
| You are | In the loop | On the review |
| It hands back | A suggestion | A branch and evidence |
Why do agents need supervision at all?
Because the failure mode of a long agent run is not bad code — it is stopping. An agent that hits a failing build, a missing dependency, or an ambiguous instruction will very often produce a fluent summary of what it *would* do next and end the turn. Nobody is watching at 2am, so the run does not resume; it just ends, and the morning's finding is that four hours produced a plan.
The second failure mode is quieter. An agent that does finish will report success, because reporting success is a text-generation task and it is very good at those. Supervision is what makes that report checkable: the standard is declared before the run, checked mechanically after it, and the answer does not come from the model that did the work.
An agent that quits at the first red test never finished anything. An agent that grades its own work never proved anything.
How does a supervisor keep a run going?
By treating a failure as an event in the run rather than the end of it. In Relay that is explicit and bounded: --autonomous continues the work until completion gates pass, with a retry count per gate, a cap on follow-ups, and a wall-clock limit — defaults of 3 retries, 12 assistant turns and 30 minutes, all of which you raise for an overnight task.
The bounds matter as much as the persistence. An agent that never stops is not supervised either; it is unattended. A budget is what makes "keep going" safe to say.
What does a supervisor verify?
Whatever you told it to, checked by something other than the model. Gates are commands: your test suite, your type checker, your linter, your own script. They pass or they do not, and the run does not end while one is failing. When the work is done, the evidence is a signed artifact rather than a paragraph — what a signed receipt proves covers what is inside one.
That is also the honest boundary. A supervisor proves the checks you wrote were run against the code you have. It does not prove the checks were the right ones. Tamper-evident, not tamper-proof.
Do I need an account to use one?
No. The local agent installs in one command, runs against your own model provider, and never checks a licence:
curl -fsSL https://relayevals.com/harness/install.sh | bash
You sign in only when you want a team-wide standard, machine visibility, or CI trust — the split is on the pricing page, and the quickstart gets a first supervised run going in a few minutes.