agent-loop

An iterative agent loop runs subprocesses against a target multiple times under hard guardrails like iteration caps and cost ceilings. It provides evaluation and iterative feedback mechanisms to converge on a target goal in automated software development cycles.

Headless agent iteration handles interruptions by saving state to a checkpoint file, allowing interrupted or capped runs to be resumed. It also performs automatic cleanup when receiving SIGINT or SIGTERM signals during the evaluation loop.

Yes, you need claude and gate.py to run automated evaluation workflows. The loop requires executing `claude -p --agent` for iteration and `gate.py --json` for post-iteration scoring and deciding whether to iterate again.

Guardrails for managing agent loops include iteration caps, cumulative cost ceilings, context-fill refusal, and oracle-score plateau detection. These hard limits prevent runaway processes during automated software evaluation and iterative testing.

Use a model-judged evaluation loop when projects require iterative testing with deterministic and model-judged scoring to converge on a target goal. It is ideal for automated software development cycles ensuring implementations reach desired quality levels.