Agents
Nine iterations of BabyAGI in three years, and yet the bit that @yoheinakajima kept coming back to was graphs. @aiDotEngineer published wher…
Nine iterations of BabyAGI in three years, and yet the bit that @yoheinakajima kept coming back to was graphs. @aiDotEngineer published where that landed, 'Active Graph Agent Runtime (BabyAGI 4)', on
Nine iterations of BabyAGI in three years, and yet the bit that @yoheinakajima kept coming back to was graphs. @aiDotEngineer published where that landed, "Active Graph Agent Runtime (BabyAGI 4)", on YouTube. The legend Yohei is Managing Partner at Untapped Capital. The talk is a working argument for building an agent around an immutable event log instead of around the LLM, with code, reference agents, and experiment results behind it. - The log is the agent. One immutable typed event log holds what the agent did and every change to the agent itself, and it projects the graph that is the agent's state. - Behaviors replace the control loop. They react to graph changes and emit events. LLMs never talk to each other, only to shared state, so replay, rollback, and forking come natively. - Policies decide what can change. Adding a research source is cheap. Editing a prompt can require a human. A new fact can require that nothing contradicts it. - Views are context management as a graph query, Handing a behavior the subset of the graph it should see. - Packs, not skills. Memory, identity, tools, secrets, chat: each bundles object types and behaviors, so you can swap one memory pack for another. - A runtime, not a harness. He rebuilds ReAct on top of it. On goal created, add a thought. On thought created, run reason. - The log doubles as memory. On LongMemEval, no fact or entity extraction, just embed the query and pull the messages around the hits. When his API key ran out at question 350, the run picked up at 353 instead of starting over. - Self-modification with gates. Regimes classifies the failure, lets the agent edit only the matching part of itself, then requires a static check, a sandbox check, and a measured rerun before a patch is accepted. Loops of 8 to 13 accepted 4 or 5, with modest but statistically significant gains. - It remembers what failed. Around 80 tuning passes on a deterministic Pokemon trading card agent for a Kaggle competition, 20 to 30 accepted, and everything that didn't work stayed on the record. - Old architecture, new workers. Blackboard and Kafka have decades of writing behind them while LLM agents have three years, which is his hypothesis for why coding agents write this style well. I'm working through the published talks from AI Engineer World's Fair sharing summaries and takeaways. Follow for more! Media
Related
- just watched @yoheinakajima's Active Graph talk at AI Engineer. not a code call graph but the agent run itself is the graph: goals, tasks, e…
- [[replit-longer-form-video-on-active-graph-7-min-22-sec---flip|@Replit longer form video on Active Graph [7 min 22 sec] - flip the agent architecture - 1970s blackboard system - rollback, fork, diff agen…]]
- Active Graph really feels like the culmination of all of my BabyAGI and graph experiments. [fyi, technical history of babyagi: http://babyag…
- i'm excited to open source Active Graph: an event-sourced reactive graph runtime for long-running, agents 🔄🧠 events/logs projects a graph.…
Source: Yohei Nakajima (X) | 2026-08-02