Evolve as a Team: Collaborative Self-Evolution for LLM-based Multi-Agent Systems

Researchers introduce Meta-Team, an experience-driven framework that enables multi-agent LLM systems to collaboratively self-evolve by learning from their own execution failures. The system coordinates post-task communication among agents to identify and implement improvements across individual behaviors, inter-agent coordination, and team-level organization, demonstrating consistent performance gains across six benchmarks.

Meta-Team addresses a fundamental challenge in deploying LLM-based multi-agent systems: the difficulty of identifying and correcting failures that emerge from complex, prolonged interactions between multiple agents. Traditional design-time approaches struggle with the combinatorial complexity of agent interactions, while post-deployment fixes remain costly and inefficient. This research shifts the paradigm toward systems that can diagnose and improve themselves through execution experience.

The framework's innovation lies in its ability to preserve individual agent contexts while enabling structured post-task collaboration. Rather than treating failure analysis as a centralized problem, Meta-Team distributes evidence gathering across agents, allowing each to contribute observations about what went wrong and why. This distributed approach mirrors how human teams conduct retrospectives, scaling the concept to artificial systems. The multi-scale evolution capability—operating simultaneously at individual agent, inter-agent coordination, and organization levels—enables comprehensive system improvement rather than localized fixes.

For developers and organizations deploying complex agentic systems, this represents meaningful progress toward more resilient and maintainable deployments. Real-world applications spanning logistics, scientific research, and autonomous planning rely on multi-agent coordination; systems that improve through experience reduce ongoing engineering overhead and decrease failure rates in production environments. The benchmark results across six long-horizon tasks demonstrate practical applicability beyond theoretical proof-of-concept.

Looking forward, the integration of experience-driven evolution into agentic frameworks will likely become standard practice. Key questions remain around scalability to larger agent teams, performance in adversarial settings, and the computational overhead of continuous self-evolution cycles. Teams building production agentic systems should monitor how these capabilities integrate into mainstream frameworks and model providers.

• →Meta-Team enables multi-agent LLM systems to autonomously improve by analyzing their own execution failures through collaborative post-task communication.
• →The framework operates across three improvement scales: individual agent behaviors, inter-agent coordination patterns, and team-level organization structures.
• →Distributed evidence gathering allows each agent to contribute observations, scaling beyond centralized failure analysis bottlenecks.
• →Benchmark testing across six long-horizon tasks shows Meta-Team consistently outperforms hand-crafted systems and prior evolution methods.
• →Experience-driven self-evolution reduces engineering overhead for maintaining complex multi-agent deployments in production environments.