TPGO: A Self-Improving Framework for Multi-Agent Systems via Textual Parameter Graph Optimization
Researchers have introduced a new framework known as Textual Parameter Graph Optimization (TPGO) to advance the autonomous development of multi-agent systems. In this setup, these systems are depicted as Textual Parameter Graphs (TPGs), with agents, tools, and workflows acting as adjustable and modular components. The framework utilizes 'textual gradients'—feedback in natural language that comes from execution logs—to pinpoint issues and suggest precise modifications. At its core is the Group Relative Agent Optimization (GRAO) method, which allows the system to refine its optimization strategies over time, overcoming the limitations of existing static optimizers that lack structural awareness or learning from experience. This work addresses the complexities of 'Agent Engineering' in creating and optimizing multi-agent systems.
Key facts
- TPGO models MAS as a Textual Parameter Graph (TPG) with modular nodes.
- Textual gradients are derived from execution traces to guide evolution.
- GRAO is a novel meta-learning strategy for self-improvement.
- Existing automatic optimization methods focus on flat prompt tuning.
- Current optimizers are static and do not learn from experience.
- TPGO enables a multi-agent system to learn to evolve.
- The framework pinpoints failures and suggests granular modifications.
- The work addresses the complexity of Agent Engineering.
Entities
Institutions
- arXiv