A Self-Reflective Multi-Agent Collaboration Framework for Dynamic Software Engineering Tasks

Large Language Model (LLM) based multi-agent systems have demonstrated remarkable potential in automating complex software engineering tasks. However, existing frameworks such as MetaGPT and AutoGen suffer from critical limitations including static role assignments, cascading hallucinations in long-horizon tasks, and the absence of experience accumulation mechanisms. We propose Eco-Evolve, a self-reflective multi-agent collaboration framework that addresses these challenges through three key innovations: (1) a dynamic topology generation mechanism that adaptively constructs agent communication graphs based on task complexity, (2) a system reflection module featuring a dedicated Critic Agent for deliberate verification at critical checkpoints, and (3) an error-driven self-evolution mechanism inspired by Hindsight Experience Replay (HER) that enables prompt optimization through experiential learning. We evaluate Eco-Evolve on SWE-bench Verified and DevBench, achieving62.3% and 73.5% respectively, representing improvements of26.6% and 14.7% over the strongest baseline. Comprehensive ablation studies validate the contribution of each component, demonstrating that integrating dynamic collaboration, deliberate reflection, and continuous evolution significantly advances the state of the art in automated software engineering.