TPGO framework lets multi-agent systems optimize themselves

Shan He, Runze Wang, and Zhuoyun Du introduce Textual Parameter Graph Optimization (TPGO), a framework that addresses two key shortcomings in current multi-agent system (MAS) optimization: a lack of structural awareness and the static nature of existing optimizers. Rather than treating a MAS as a flat collection of prompts, TPGO represents it as a Textual Parameter Graph (TPG), in which agents, tools, and workflows are modular nodes that can each be independently targeted for optimization. To identify where a system is going wrong, TPGO derives "textual gradients" — structured natural language feedback extracted from execution traces — that pinpoint specific failures and suggest granular modifications at the node level.

The central innovation is Group Relative Agent Optimization (GRAO), a meta-learning strategy that accumulates experience across optimization rounds.

The central innovation is Group Relative Agent Optimization (GRAO), a meta-learning strategy that accumulates experience across optimization rounds. By analyzing a history of past successes and failures, GRAO becomes progressively better at proposing effective updates, enabling the system to learn how to optimize itself rather than relying on a fixed optimization policy. The authors validate TPGO on complex benchmarks including GAIA and MCP-Universe, reporting that it significantly enhances the performance of state-of-the-art agent frameworks through fully automated, self-improving optimization — without requiring manual re-engineering of the underlying system.

Key facts

1. 01TPGO (Textual Parameter Graph Optimization) is a new framework for automated, self-improving optimization of multi-agent systems (MAS).
2. 02It models a MAS as a Textual Parameter Graph (TPG), where agents, tools, and workflows are modular, optimizable nodes.
3. 03"Textual gradients" — structured natural language feedback derived from execution traces — are used to pinpoint failures and suggest targeted fixes.
4. 04Group Relative Agent Optimization (GRAO) is a meta-learning strategy that learns from historical optimization successes and failures to improve future updates.
5. 05GRAO enables the system to progressively improve its own optimization strategy rather than relying on a static optimizer.
6. 06Experiments on the GAIA and MCP-Universe benchmarks show TPGO significantly improves state-of-the-art agent frameworks.
7. 07The paper is authored by Shan He, Runze Wang, and Zhuoyun Du.