EvoTeam: Multi-Agent Collaboration via Evolving Teams

EvoTeam is a collaboration-centric evolution framework that maintains reusable agent teams and jointly evolves team organization, team-specific interaction rules, and individual agent skills to improve LLM multi-agent systems across heterogeneous tasks.

Existing MAS evolution often improves individual agents, but better individual agents do not necessarily produce better collaborative teams. Recent collaboration-level evolution still constrains team reorganization within a shared structure and reuses interaction rules across incompatible team contexts. The paper's preliminary study shows that continued evolution across heterogeneous tasks can reduce target-task performance relative to target-only evolution.

EvoTeam maintains a global agent pool and a library of reusable agent teams, where each team has a description and candidate agent roster. For each task, it selects a compatible team or creates a provisional team, then a Chairman recruits an execution team from the roster and assigns subtasks. Team-specific interaction rule sets are injected as soft guidance, and agents select applicable rules based on coordination needs. After execution, three reflectors update the team organization, its interaction rules, and individual agent skills. Interaction rules remain team-specific, while agent skills remain reusable across team contexts.

• Under mixed-benchmark evolution, EvoTeam ranks first on all six metrics and achieves an average score of 59.58, surpassing Meta-Team by 4.03 points. §4.2
• Under individual-benchmark evolution, EvoTeam achieves the highest average score of 58.85 and leads on four of the six evaluation metrics, with a GAIA margin of 78.00 versus 73.00. §4.2
• In cross-backbone evaluation, EvoTeam outperforms MAS in all six configurations; on GAIA, gains reach 5.00 and 18.00 points for GPT-5.6-Luna and DeepSeek-V4.1-Flash respectively. §4.3
• Cross-backbone transfer remains within 4.00 points of same-backbone performance across all six paired comparisons; on GAIA it achieves 77.00 versus 78.00 under same-backbone evolution. §4.3
• Ablation shows learned interaction rules add 10.00 and 8.75 points on GAIA and LOCA, while task-conditioned teams without interaction rules add 6.00 and 2.50 points on GAIA and LOCA. §4.4
• On evaluated cost-efficiency tasks, EvoTeam achieves the highest score while costing less per task than Meta-Team and ReCreate; Skill-MAS has lower cost but lower performance. §4.5

• The paper states that its evaluation does not systematically examine collaborative evolution in multimodal interactive environments.
• The paper states that EvoTeam accumulates experience through explicit updates while keeping LLM parameters fixed, so effective use depends on the base model correctly interpreting and applying stored guidance.
• Experiments leave open broader backbone generalization: cross-backbone transfer is evaluated with DeepSeek-V4.1-Flash and GPT-5.6-Luna only.

• Preliminary study: both Inter-CT and Intra-CT evolution settings underperform target-only Meta-Team on all three metrics. §2.1
• The paper identifies constrained team reorganization and inappropriate reuse of interaction rules as potential failure modes. §2.2
• EvoTeam maintains a global agent pool and reusable team library; each team has a description and candidate roster, and an execution team is recruited per task. §3.1
• Team-specific interaction rule sets narrow reuse scope, and the selected agent team determines the available rule set. §3.2
• Multi-level evolution uses reflectors to update team descriptions or rosters, interaction rules, and individual agent skills; rules stay team-specific while skills are reusable across teams. §3.3
• Across four benchmarks and six task categories, EvoTeam leads all six metrics under mixed-benchmark evolution, averaging 59.58 versus 55.55 for the strongest baseline. §6
• Ablations show value from both task-conditioned team organization and learned interaction rules. §4.4
• Cost comparison places EvoTeam at the strongest performance among compared methods with competitive per-task cost. §4.5