Observability · ~7 min

The Log Is the Truth

Stop letting agents write files directly. Have them emit intentions to an append-only log; a deterministic orchestrator applies the effects — and you can replay the whole thing.

Long-horizon agents fail two ways: context degradation (early decisions fall out of the window, so the agent repeats or contradicts itself) and non-deterministic mutation (the agent writes files directly, so you can't audit, replay, or verify that intention matched effect). ESAA — Event Sourcing for Autonomous Agents — separates the two.

1 Cognition emits; the orchestrator mutates

The agent — the cognitive layer — produces a structured JSON intention. It never writes a file. A deterministic orchestrator validates the schema, appends the event to an append-only log, then applies the effect.

2 Replay re-derives state from scratch

Because the log is append-only and every effect is deterministic, you can replay activity.jsonl and rebuild project state from nothing. If the derived state matches the filesystem, the execution record is verified.

This gives three properties at once: forensic traceability (every change explained by a logged event with timestamp and agent identity), immutability (no in-place updates, so history can't be silently revised), and reproducibility (same event sequence → same final state). The log is the source of truth; current state is a derived projection.

3 Concurrent agents, no write conflicts

In the paper's clinical-dashboard case study, 4 concurrent heterogeneous LLMs worked different tasks from one roadmap. The orchestrator serializes event persistence and state mutation, so they never collide — and they need no coordination logic of their own. They stay cognitively independent: emit an intention, receive an updated roadmap.json view.