Procedural Graphs: Self-Evolving Execution Structures for LLM Agents

Imagine you're building an AI assistant that needs to manage complex tasks over time, like scheduling meetings or coordinating projects. The challenge is that as tasks become more intricate, agents often lose track of what they should do next, leading to mistakes like using the wrong tool or repeating actions that don't help. This is what's called losing context or procedural drift, which can severely hinder performance in real-world applications. Current methods typically rely on simple memory systems that don't effectively guide agents through these long sequences of actions. They often lack a structured way to represent the knowledge of what to do and when to do it.

To address this, the authors propose the Procedural Graph, a framework that organizes procedural knowledge into a graph structure. This allows agents to reference a network of procedures, helping them make informed decisions at each step. The graph evolves as the agent learns from both successful and failed attempts, refining its structure to improve future performance. This self-evolving aspect means that agents can adapt their strategies without needing constant manual updates. Compared to traditional memory-based approaches, the Procedural Graph consistently outperforms them, providing a more robust solution for building agents that can navigate complex tasks effectively.

The introduction of the Procedural Graph represents a significant advancement in organizing procedural knowledge for agents.

The approach shows consistent performance improvements across various datasets, though further validation could strengthen claims.