A tamper-evident, hash-chained record of what each agent actually did — real autonomy, not an LLM call wrapped in a workflow. The chain is independently re-verified on every load.