Notes on AI agent safety, guardrails, and LLM operations — explained simply.
The agent edits its own Python (the scaffold), not the model's weights. The one detail that matters: it samples a parent from a growing archive each round and keeps improvements as new branches — it never overwrites the current best. That archive is what lets weak ancestors seed later breakthroughs.