Agents You Can Leave Running

Leave it running overnight. Trust it by morning.

You shipped the loop. Reason, act, observe. It works in the demo. Then you let it run while you sleep, and you wake up to forty pull requests, a context window full of confident nonsense, an agent that graded its own broken work an A, and a token bill that made your CFO call a meeting. The hard part of agentic engineering was never the loop. It was learning to trust one.

The reason-act-observe cycle is decades old and the easy part. What consumes the real engineering hours is the outer control system that drives it, and Agents You Can Leave Running is the advanced field manual for that system: prove the work, stop the runaway, remember across resets.

The control-system playbook for long-running autonomous agents:

• The ungameable check replaces the self-verifying agent grading its own exam with verification and evals the loop cannot influence: a fresh model's review plus deterministic gates.
• Halt conditions that hold wire in no-progress detectors, dollar ceilings, and circuit breakers: guardrails that stop runaway cost and goal drift before a run touches money.
• Memory across resets moves state out of the context window into artifacts the next turn can reload, so the loop picks up exactly where it left off.
• Observability as the control layer uses tracing to close the loop, so every overnight run leaves evidence you can read instead of a mystery you have to replay.

Read it and you will stop babysitting your agents turn by turn, know which checks a loop can game and which it cannot, place the ceilings before a run gets expensive, and hand an agent a night's work you can trust by morning.

For AI engineers and applied researchers who already know the ReAct loop and are building production agentic systems: this is the handbook for what comes after it. Part of the Build Agents You Can Trust series, in The Verifier's Library.