Meta-Team
The reference implementation of a research paper on collaborative self-evolution for LLM-based multi-agent systems, where a team of agents reflects together after each task at the level of individual agents, their interactions, and the team's own organization.
Meta-Team is built around the idea that a multi-agent system should evolve as a team, not only execute as one. After each task, the system reflects at three levels: an agent level, where prompt patches and new skills are added to an individual agent's scaffold; an interaction level, where teammate profiles and collaboration notes update how agents communicate; and a team level, where the shared constitution, organization and coordination rules themselves change. Preserving each agent's execution context and coordinating post-task communication lets agents exchange distributed evidence that drives this multi-scale evolution.
The repository is configured to run against any OpenAI-compatible LLM API, with published experiments using Claude Sonnet 4.6, and evaluates on coding and agentic benchmarks including SWE-bench Pro, BeyondSWE, LOCA-Bench, GAIA and LoCoBench, some of which require Docker images for the benchmark tasks.
It is a Python research implementation for the paper 'Evolve as a Team: Collaborative Self-Evolution for LLM-based Multi-Agent Systems,' requiring Python 3.10 or later.