The Log is the Agent — Yohei Nakajima's ActiveGraph Event-Sourced Agent Architecture
Yohei Nakajima (creator of BabyAGI) introduces ActiveGraph — an architecture where the append-only event log is the source of truth and the working graph is its deterministic projection.
Introduction: Flipping the Agent Stack
Yohei Nakajima, creator of BabyAGI, has published a paper proposing ActiveGraph — an architecture that inverts how AI agents are built. Instead of building around the LLM and bolting logging on later, ActiveGraph makes the append-only event log the source of truth.
The paper, "The Log is the Agent: Event-Sourced Reactive Graphs," describes a runtime where the working graph is a deterministic projection of the event log. Components react to graph changes and emit new events — no component directly instructs another.
Three Killer Properties
Deterministic replay: Any run can be perfectly replayed from its event log. Cheap forking: Branch a run at any point without re-executing the shared prefix — like git for agents. End-to-end lineage: Trace from high-level goal down to the individual model call that produced each artifact.
Why This Matters
Current agent memory systems rely on retrieval + summarization, which is lossy. ActiveGraph provides a lossless audit trail. This substrate is ideal for self-improving agents that need to audit their own decisions.
From BabyAGI to ActiveGraph
Nakajima extends the BabyAGI lineage and prior graph-memory research. The implementation is Apache-2.0 open source with a quickstart demo, deterministic replay, fork-and-diff, and lineage tracing tools.
Conclusion
ActiveGraph could enable agent development workflows similar to software engineering — branch, diff, merge, and review agents the way we do code. This is a glimpse of where agent infrastructure is heading.