Procedural Graphs: Self-Improving LLM Agent Execution Structures

Procedural Graphs: Self-Evolving Execution Structures for LLM Agents When AI Agents Start Writing Their Own "Brain Circuits" Published: September 10, 2026 | Reading time: 12 minutes The Revolutionary Research On September 9, 2026, researchers Yuxing Lu, Yicheng Chen, and Shanchan Wu published a groundbreaking paper on Procedural Graphs — a self-evolving execution structure for LLM agents that can literally rewrite its own "brain circuits." Paper: arXiv:2609.08593 The Problem with Current LLM Agents Today's LLM agents (like AutoGPT, LangChain agents) work like this: Maintain a growing memory of past thoughts, observations, actions Generate next action based on this memory No fixed structure, no predefined流程 This works for simple tasks, but breaks down for complex ones: Lost goals: Forget what they're supposed to do in long interactions Action mismatch: Call tools in wrong order (e.g., analyze before searching) Repeated labor: Try same ineffective operations repeatedly No planning: No global view, step-by-step navigation Analogy: Like a chef without a recipe — overwhelmed by complex dishes. The Solution: Procedural Graphs Procedural Graphs organize procedural knowledge ("how to do") just like Knowledge Graphs organize factual knowledge ("what is"): Type Structure Question Knowledge Graph (Entity, Relation, Entity) What is? Procedural Graph (Procedure, Relation, Procedure) How to? Core Components Nodes (程序步骤): Each node represents a procedural step Contains: description, expected I/O, success/failure conditions Like a box in a flowchart Edges (关系): Represent relationships between steps Types: Sequential ("then"), Conditional ("if...then"), Parallel ("simultaneously") Like arrows in a flowchart Attributes (属性): Execution probability, average time, success rate, common error patterns Update with execution experience How It Works: A Concrete Example Task: "Book a cheap flight from Beijing to Shanghai, departing tomorrow." Procedural Graph: [Start] ↓ [Search Flight Info] ↓ [Filter Low-Cost Options] ↓ [Check Seat Availability] ↓ [If Available: Fill Passenger Info] ↓ [Select Payment Method] ↓ [Confirm Order] ↓ [End] ↓ [If Unavailable: Return to Filter] ↓ [If No Satisfying Result: Expand Search] ↓ [Search Nearby Airports] ↓ [Return to Filter] This is a directed graph with branches, loops, and conditionals — far more powerful than linear memory. The Self-Evolution Mechanism The most amazing capability is self-evolution: Evolution Cycle Collect trajectories: Record complete execution paths Compare analysis: Contrast failed trajectories with successful ones Identify differences: Find where things went wrong Generate edits: LLM Refiner proposes modifications Verify and retain: Test on validation set Three Types of Edits Topology Edits: Add new nodes (new steps) Remove redundant nodes Add/modify edges (change flow structure) Attribute Edits: Update node success rate statistics Adjust condition thresholds Update execution probabilities Content Edits: Modify node descriptions for accuracy Update guidance language for effectiveness From Skeleton to Maturity Researchers tested three initialization methods: Initialization Evolution Speed Final Performance Empty Graph (only "Start" node) Slower Close to others Minimal Skeleton (basic manual nodes) Fastest Best Expert Prior (human-designed) N/A Repairable if flawed Amazing Discovery: Even starting from a flawed expert prior, the evolution mechanism can "repair" it to achieve good performance. This shows robustness — doesn't require perfect initial design. Why Graphs Beat Memory Dimension Pure Memory Workflow Memory Procedural Graph Structure None Case-level abstract Procedure-level Generalization Poor Medium Good Explainability Poor Medium Good Evolution None Limited Strong Efficiency Low Medium High Key Advantage: Procedural Graphs abstract the general flow for a class of tasks, not just specific past cases. Analogy: Workflow Memory = Remember "Last time I made Mapo Tofu, I stir-fried meat first, then added bean paste" Procedural Graph = Understand "General stir-fry flow: Heat pan → Add oil → Stir-fry main ingredient → Season → Serve" The latter generalizes to any stir-fry; the latter can only repeat Mapo Tofu. Experimental Results WebShop (Web Shopping) Task: Purchase items on e-commerce sites based on natural language instructions Graph vs Memory: 15-25% success rate improvement Evolved Graph vs Initial: 10-20% improvement ALFWorld (Home Tasks) Task: Execute daily tasks in simulated home environment Graph helps remember complex object interaction sequences HotPotQA (Multi-hop QA) Task: Multi-step information retrieval and reasoning Graph optimizes retrieval strategy and evidence integration Tool Use Tasks Task: Combine multiple APIs to complete complex goals Graph ensures correct tool call order and parameter settings Key Findings Finding 1: Cross-LLM Generalization Graph evolved on one LLM (e.g., GPT-4) can transfer to another (e.g., Claude or Llama) Shows graphs capture task structure, not model-specific traits Finding 2: Few-Shot Advantage Effective even with few examples (