Feed
Procedural Graphs: Self-Evolving Execution Structures for LLM Agents
A graph of what-to-do-next steps guides AI agents and rewrites itself from their failures.
85.0%
Test survival in financial simulator
Unguided agent never survived the full 132-month horizon
How it works
Agents lose the plot
Over long tool-using tasks, agents forget goals, call tools out of order, and repeat useless actions.
Graph of procedures
Store steps as (procedure, relation, procedure) edges. Each step, read the local subgraph and turn it into short advice.
Self-edited, consistently better
An LLM refiner edits the graph from failures, keeping edits that pass validation. It wins 19 of 24 head-to-heads.
From the paper