cleanexpo Agentic Loop
Two-prompt system: task prompt + stop guard. Agent works -> tries to stop -> guard checks criteria -> not met -> continues.
Safety Rails
- max_iterations: 20
- max_tokens: 200000
- max_runtime_minutes: 60
- Detect oscillation (fix A breaks B) after 3 iterations
Overnight-loop guardrails (Ralph Wiggum, Pocock 2026-07)
- Iteration cap:
max_iterationsabove is the backstop against never terminating — keep it. - Append-only progress file: each iteration appends learnings ("append", never "update") — memory across context resets.
- One small feature per iteration: never batch ("do as many as you fancy" recreates the context bloat the loop exists to avoid).
- Green-CI gate: typecheck + tests pass inside the iteration before its commit; broken commits blind the next fresh context.
- Completion sentinel: prompt outputs an explicit token (e.g.
PROMISE COMPLETE); the loop greps it and exits early rather than relying on the stop guard alone.
[[pocock-ralph-wiggum-overnight-loops-2026-07-14-ingest]]