program.md contains a directive that tells your agent to never pause, never ask "should I continue?", and never suggest stopping points. It must run indefinitely until you manually kill it.
Without this directive, LLMs default to polite, interactive behavior. They summarize progress, ask for confirmation, or suggest good stopping points. Any of these halts the autonomous loop.
This is why LLM selection matters. Claude Opus follows the directive and has sustained + hour runs of + experiments. Codex ignores it. Karpathy filed Issue # documenting that Codex halts mid-experiment, breaking the loop.