Session brief

Active Graph Agent Runtime (BabyAGI 4) — Yohei Nakajima, Untapped Capital

Yohei Nakajima , Untapped Capital

Overview

This talk introduces ActiveGraph, an AI agent runtime built around an immutable event log. Instead of bolting memory and logging onto an LLM, ActiveGraph uses the log as the core of the agent's state. This approach enables features like replays, rollbacks, and forks by default, as every action and state change is recorded. The system then uses behaviors and policies to react to changes in the projected graph, which represents the agent's state.

Who should watch

• AI engineers building agentic systems
• Developers interested in agent state management and persistence
• Builders exploring alternative architectures for AI agents beyond LLM-centric designs
• Those facing challenges with agent reliability and state recovery during long-running tasks

Key takeaways

• ActiveGraph is built around an immutable event log, where every action and state change is recorded, forming the agent's state.
• This log-centric design inherently provides capabilities such as replays, rollbacks, and forks.
• Behaviors react to changes in the graph projected from the log, while policies govern what the agent can modify.
• The architecture draws inspiration from 1970s blackboard systems and Kafka, with LLMs interacting through shared state rather than direct communication.
• The system supports self-improvement loops where agents can propose, sandbox, and integrate patches that demonstrably increase accuracy.
• An "ActiveGraph Lab" component allows agents to run their own experiments and even identify and fix bugs in their own code.
• The talk suggests that AI might architect systems in this log-based style more effectively due to extensive training data on such patterns.
• An experiential world model is considered essential for an agent to effectively operate.

Notable quotes

Watch on YouTube →

Unofficial community note. Prefer the recording for nuance.