Yohei Nakajima Open Sources ActiveGraph Runtime For Reactive AI Agents
current agent systems coordinate through conversations and workflows. Active Graph explores what happens when agents coordinate through evolving shared state instead
this proposal suggests that long-running agents need a proper state layer with: types, persistent, reactive, replayable, forkable, inspectable state
Yohei@yoheinakajimai'm excited to open source Active Graph: an event-sourced reactive graph runtime for long-running, agents 🔄🧠 events/logs projects a graph. reactive behaviors react and affect the graph. fork-and-diff agent runs. no A2A, no workflows, no DAG site: http://activegraph.ai docs: http://docs.activegraph.ai github: https://github.com/yoheinakajima/activegraph quick start: pip install activegraph this is an early experiment in a new paradigm for agent architecture 🧪
the core concept is a graph that represents everything about the agents knowledge, history, behaviors, capabilities
graph is made of events behaviors react to graph changes relationships can carry behaviors patch & propose to edit graph views are scoped view of graph frames are bounded context for a run policies set rules
Yohei@yoheinakajimacurrent agent systems coordinate through conversations and workflows. Active Graph explores what happens when agents coordinate through evolving shared state instead this proposal suggests that long-running agents need a proper state layer with: types, persistent, reactive, replayable, forkable, inspectable state
this is how you add an event, fork and cache a run, and then find the diff between a parent and fork
in this example, the fork shares the parent's event log up to event 142. from 143 onward it diverges
being able to do this should help with building self-improving agents
Yohei@yoheinakajimathe core concept is a graph that represents everything about the agents knowledge, history, behaviors, capabilities graph is made of events behaviors react to graph changes relationships can carry behaviors patch & propose to edit graph views are scoped view of graph frames are bounded context for a run policies set rules
relations can carry behaviors
- depends_on can unblock work - contradicts can know how to trigger review - derived_from can know how to invalidate downstream artifacts
Yohei@yoheinakajimabehaviors are small reactive units > state changes, behaviors react, behaviors emit more state changes no central orchestrator
this is what a ReAct agent looks like using ActiveGraph
No explicit ReAct while-loop.
thought → action → observation → thought
Each step is just an object landing on the graph. The graph is the agent state. The event log is the reasoning trace.
Yohei@yoheinakajimathis is what babyagi looks like using activegraph the graph is the loop (this is simplified code, you can find a real example in examples/babyagi.py)
#1234Yohei@YOHEINAKAJIMAif you're curious to read more about this project, here's an article i shared yesterday explaining the background
Yohei@yoheinakajimahttp://x.com/i/article/2056734659736985600
#1234Yohei@YOHEINAKAJIMAand the article before talked even more broadly about the state of stateful agents, which inspired this project
Yohei@yoheinakajimahttp://x.com/i/article/2056590933874147328
Active Graph really feels like the culmination of all of my BabyAGI and graph experiments.
[fyi, technical history of babyagi: http://babyagi.wiki]
would love to hear thoughts if you try it out!
the easiest way to test is point your favorite coding agent to "http://docs.activegraph.ai" and asking it to build something or ask it questions
Yohei@yoheinakajimaand the article before talked even more broadly about the state of stateful agents, which inspired this project
@grok @perplexity_ai review http://docs.activegraph.ai, analyze the approach critically, and explain in detail where this sits in the evolving landscape of LLM based agent architecture
Yohei@yoheinakajimai'm excited to open source Active Graph: an event-sourced reactive graph runtime for long-running, agents 🔄🧠 events/logs projects a graph. reactive behaviors react and affect the graph. fork-and-diff agent runs. no A2A, no workflows, no DAG site: http://activegraph.ai docs: http://docs.activegraph.ai github: https://github.com/yoheinakajima/activegraph quick start: pip install activegraph this is an early experiment in a new paradigm for agent architecture 🧪
for anyone curious, this was the result of many experiments bouncing around but this version uses...
initial ideation: chatgpt for ideas/@replit for quick build final repo build: opus 4.7 prompting claude code pypi testing: ran so many tests w @replit who wrote reports that I fed back into claude launch video: @elevenlabs and claude
Yohei@yoheinakajimaActive Graph really feels like the culmination of all of my BabyAGI and graph experiments. [fyi, technical history of babyagi: http://babyagi.wiki] would love to hear thoughts if you try it out! the easiest way to test is point your favorite coding agent to "http://docs.activegraph.ai" and asking it to build something or ask it questions
#1234Yohei@YOHEINAKAJIMA@Replit narrative walkthrough video of repo for anyone interested: