← back
Active Graph Agent Runtime (BabyAGI 4) — Yohei Nakajima, Untapped Capital
Takeaway
Making the event log the agent’s source of truth enables auditable execution and controlled self-modification.
Summary
- ActiveGraph is an experimental event-sourced runtime where an immutable log records both agent actions and changes to the agent itself.
- Graph state is projected from events; deterministic or LLM-powered behaviors observe changes and emit further events instead of messaging each other directly.
- Policies gate sensitive changes, such as prompt edits or conflicting facts, while graph-query views provide scoped context.
- The shared log supports replay, rollback, and forks, and reusable packs combine schemas, tools, behaviors, and policies.
babyagievent-sourcingagent-runtime
Original description
Yohei Nakajima was running a 500 question eval when his API key died at question 350. Normally that means restarting the whole long agent from scratch. Instead it rolled back one step and resumed at 353, because in ActiveGraph the log is the agent. Most people build agents around the LLM and bolt on memory and logging; Nakajima, the creator of BabyAGI, flips it and builds around an immutable event log. Every action and every change to the agent flattens into one typed log, which projects a graph that is the agent's state, so you get replays, rollbacks, and forks for free. On top of the log sit behaviors that react to graph changes and emit events, policies that decide what the agent may change on its own versus what needs a human or a contradiction check, and swappable packs for memory, tools, and chat. The LLMs never talk to each other; they only touch shared state, an idea he borrows from 1970s blackboard systems and Kafka, and his hunch is that AI writes this style better than modern agent code because it has decades of training data on it. The payoff is self improvement: a loop that forks the agent, proposes a patch, gates it behind sandbox tests, and keeps it only if accuracy actually rises, and a lab that reads his blog posts, runs its own experiments, and once found a bug in its own code and opened the PR. Speaker info: https://github.com/yoheinakajima/acti... https://x.com/yoheinakajima / yoheinakajima Timestamps: 0:00 - ActiveGraph and three years of BabyAGI 1:55 - Build around the log, not the LLM 3:24 - Behaviors, policies, and views 6:43 - Packs and the blackboard architecture lineage 8:12 - The log as memory, and the API key that resumed itself 9:53 - Reference agents built natively on the log 11:11 - Self improvement: regimes and controlled self modification 12:25 - ActiveGraph Lab writes its own experiments 13:02 - A Pokemon card competition as a testbed 14:33 - Surprises: why AI architects this better 15:49 - Why an agent needs an experiential world model