Modeling Multi-Agent Systems as Graphs for Auto-Optimization
leanxbt · x · 2026-07-18
This paper presents a novel approach to managing and optimizing multi-agent systems (Agent Swarms) by abstracting the agent system as a computational graph:
- Graph Structure Abstraction: Represents LLM agents as a computational graph, where nodes are operations that process data or query models, and edges define the information flow between them.
- Recursive Composition: Multiple graphs can be recursively assembled into larger composite graphs, turning complex agent swarms into a unified object rather than a collection of scattered scripts.
- Bi-Level Optimization: The system features two optimizers. The first automatically fine-tunes prompts for individual nodes; the second uses reinforcement learning (RL) to dynamically alter the graph's connections, optimizing the overall collaborative architecture of the agents.