Yohei Nakajima, creator of BabyAGI and a venture capitalist at Untapped Capital, is making a provocative bet: the future of autonomous AI agents is not about building smarter models, but about giving agents an unforgeable diary. In a recent interview, he detailed ActiveGraph, an open-source runtime that treats an immutable event log — not the LLM — as the core unit of agent design. This log-as-source-of-truth architecture enables agents to natively replay, roll back, and fork their own state, remembering every failed experiment alongside successes. Nakajima revealed concrete results: a Pokémon trading card deck was iteratively optimized through 80 self-modification cycles, with each change validated across 200 simulated battles, reaching a 27% win rate. A separate self-improvement loop called Regimes achieved statistically significant benchmark gains by forking, sandbox-testing, and selectively accepting its own code patches. Nakajima argues that as models improve, the agent harness becomes more critical, not less — and that long-running agents need an “experiential world model” derived from their own lived history. The architecture, which he admits is unintuitive for humans to write but easily handled by AI code generation, points toward a future where agents not only execute tasks but build identity from their accumulated experience.
