2026-10-09 校核：论文 的重放依赖确定性契约和已记录的模型/工具响应；重放已记录的事件可重建图，不代表再次向模型或外部服务发同样请求就必然得到原结果。追加 timeline 本身也不构成完整 event sourcing 或全链路重放；Hermes/gbrain 对照是架构建议，未做本地实现验收。
The Log is the Agent: Event-Sourced Reactive Graphs
Paper: arXiv:2605.21997 (May 2026)
Author: Yohei Nakajima (BabyAGI creator)
HN: 14 points, front page Jul 5 2026
Core Idea
ActiveGraph inverts the traditional agent architecture. Instead of:
LLM loop → tools → rules → logging (bolt-on)
It does:
Append-only event log → deterministic graph projection → reactive behaviors → new events
No component instructs another directly — coordination happens entirely through the shared graph.
Three Unique Properties
- Deterministic replay — any run can be fully replayed from its log
- Cheap forking — branch a run at any event without re-executing the shared prefix (like git branch)
- Full lineage — from high-level goal down to individual model call that produced each artifact
Why This Matters for Multi-Agent Orchestration
Most multi-agent frameworks struggle with:
- Observability: who did what, when, and why
- Determinism: can the same inputs reproduce the same outputs
- Forking/merging: branching a conversation or agent run for experimentation
ActiveGraph solves all three through one architectural choice: the log is the source of truth, not the model state.
Behaviors can be:
- Ordinary functions
- Classes
- LLM-backed routines
- Logic attached to typed edges
All react to graph changes and emit new events. The graph is a deterministic projection of the log — replay produces the exact same sequence.
Relevance to Hermes
This design is philosophically aligned with:
- gbrain’s append-only timeline entries (we already do event-sourced persistence)
- AMB message bus (event-driven coordination)
- Session replay (Hermes already supports session replay)
The paper explicitly extends the BabyAGI lineage and prior graph-memory research, which maps to our multi-agent patterns research in research/2026-multi-agent-orchestration-patterns.