Event Sourcing records changes as ordered facts instead of treating current state as the only truth. Applied to agents, it solves a recurring problem: separating what a model suggested from what the system actually admitted.
Intent, event and projection
An intent is the agent’s structured proposal. The Orchestrator validates it against contracts, state and boundaries. Only then is an event appended to the canonical record. Roadmap, issues and lessons are projections derived from that history.
This separation prevents a reading interface from accidentally becoming a writing authority. If a projection is corrupted, it can be rebuilt. If the record is intact, history remains available.
Transitions, not labels
A task does not “become done” because someone changed a field. It moves through admitted transitions: todo → in_progress → review → done. Every step declares an actor, prior state and evidence.
The value of replay
Replay is more than recovery. It is an architectural test: the same history must produce the same state. When it does, dashboards and automation no longer depend on free-form interpretations of a conversation.
Event Sourcing does not make the model deterministic. It creates a deterministic boundary around its actions, exactly where autonomous systems need it most.