Meta-Team lets LLM agent teams self-evolve from execution failures
Multi-agent systems that learn by sharing context and improving coordination after tasks
LLM-based multi-agent systems (MAS) excel at complex, long-horizon tasks but often fail in real-world execution, and such failures are hard to fix at design time. To solve this, researchers propose Meta-Team, an experience-driven evolution framework that lets a MAS improve based on its own execution logs. The key innovation is preserving the full execution context of each agent and enabling structured post-task communication, so agents can share distributed evidence about what went wrong. Meta-Team then performs multi-scale self-evolution: it refines individual agent behaviors, improves inter-agent coordination, and even adjusts team-level organization, turning raw execution history into reusable improvements.
Tested on six long-horizon agent benchmarks, Meta-Team consistently beats single-agent systems, hand-crafted MAS designs, and prior MAS evolution techniques. The framework enables more reliable and scalable self-improvement, suggesting that future agent teams can continuously learn from their mistakes without manual redesign. This work moves beyond static agent orchestration toward truly adaptive, self-improving multi-agent systems.
- Meta-Team preserves execution context per agent and enables post-task communication for distributed evidence exchange
- Conducts multi-scale self-evolution: improves individual behaviors, inter-agent coordination, and team organization
- Outperforms single-agent, hand-crafted MAS, and prior evolution methods across six long-horizon benchmarks
Why It Matters
Self-evolving agent teams that learn from mistakes could radically improve autonomous systems for complex real-world tasks.