Reinforced Collaboration in Multi-Agent Flow Networks

2026-05-12

AI Summary

This work addresses the challenge of error propagation in multi-agent systems, which often stems from suboptimal workflow design or erroneous individual outputs, thereby degrading overall performance. To mitigate this issue, the authors propose MANGO, a novel framework that uniquely integrates data-driven flow networks with reinforcement learning and textual gradients to jointly optimize both the collaborative structure and agent behaviors. MANGO further incorporates a skip mechanism to bypass redundant computations, enhancing efficiency without compromising accuracy. Evaluated across seven benchmark tasks, the framework achieves performance gains of up to 12.8% and improves inference efficiency by 47.4%. Notably, it also demonstrates strong generalization capabilities in unseen domains, highlighting its robustness and adaptability.

Abstract

Multi-agent systems provide a powerful way to extend large language models (LLMs) by decomposing a complex task into specialized subtasks handled by different agents. However, their performance is often hindered by error propagation, arising from suboptimal workflow design or inaccurate agent outputs, which can propagate through the agent collaboration process and degrade final results. To address the challenges, we present MANGO (Multi-Agent Network Gradient Optimization), a data-driven framework that organizes and refines agent collaboration via a flow network constructed from past successful workflows. MANGO integrates reinforcement learning and textual gradients to jointly optimize workflow paths and agent behaviors, while a skipping mechanism prevents redundant updates to well-optimized agents for improving efficiency. Extensive experiments on seven benchmarks show that MANGO achieves up to 12.8% performance improvement over state-of-the-art baselines, enhances efficiency by 47.4%, and generalizes effectively to unseen domains. Our code and datasets are publicly available at https://github.com/openJiuwen-ai/agent-store/tree/main/community/mango.

Problem

• multi-agent systems
• error propagation
• workflow optimization
• agent collaboration
• large language models

Innovation

• multi-agent systems
• reinforcement learning
• textual gradients
• flow network
• workflow optimization