EvoTeam: Continuous Multi-Agent Team Evolution Across Skills, Teams, and Interaction Rules

Continuous Work Does Not Equal Continuous Evolution

Research tasks often involve problem analysis, solution design, code implementation, experimental validation, and result integration. A single agent handling all stages must frequently switch roles and manage large contexts; multi-agent systems address this through specialized division of labor, parallel execution, and mutual verification. However, merely having multiple agents does not guarantee an efficient team. Fixed roles that cannot adapt to tasks, member capabilities that do not accumulate, and recurring handoff omissions cause organizational costs to grow with collaboration scale. Therefore, the next step for multi-agent systems is not only making each agent stronger but also enabling the team organization and collaboration mechanisms to evolve continuously.

EvoTeam Evolves Three Interconnected Levels

EvoTeam decomposes system growth into three mutually reinforcing levels:

Agent Skills focus on individual professional capabilities. Agents accumulate reusable skills from historical trajectories and task feedback, continuously improving execution quality within their functional roles.

Agent Teams focus on team collaboration architecture. As task demands and member capabilities change, the system adjusts candidate members and responsibility assignments so that suitable agents form suitable teams.

Interaction Rules focus on information handoffs. Teams learn from successes and failures to specify who should receive what information, what evidence must be provided, and what verification is required, preventing the same collaboration issues from recurring.

These levels do not occur strictly sequentially. Individual capability growth drives role adjustments, team changes create new handoff needs, and interaction problems feed back into skill updates. Individual growth thereby transforms into growth of the entire collaboration system.

Teams Become Carriers of Experience Accumulation and Reuse

EvoTeam maintains a global Agent Pool and a continuously updated Team Library, forming a closed loop through three phases:

Adaptive Agent Team Composition : For a new task, the system analyzes required professional roles and collaboration requirements, then selects or assembles a candidate team from the Team Library.

Team-Scoped Collaboration : During execution, the team's accumulated Interaction Rules guide information exchange and task handoffs. Rules are maintained per team, reducing inappropriate cross-team transfer.

Multi-Level Experience Evolution : After task completion, the system updates individual Skills, reflects on team composition, and creates, revises, or retains interaction rules based on trajectories and feedback.

Skills follow agents and can be reused across teams; member configurations and Interaction Rules belong to specific teams and accumulate continuously on tasks with compatible collaboration needs. Teams cease to be one-off orchestrations and become reusable, evolvable experience carriers.

Cross-Benchmark Growth: Teams Continue to Improve Across Domains

The paper illustrates this with a case starting from LoCoBench's wallet address verification refactoring. The system forms a team of Context Analyst, Implementer, and Verifier responsible for dependency analysis, code modification, and independent verification. The same team is then reused on BeyondSWE tasks such as interface adaptation, exception handling, and other refactoring tasks, continuously accumulating rules for interface compatibility, behavior preservation, and result verification. After the third task, an Integrator role is added to the candidate pool to support broader cross-component integration. Past tasks are no longer mere history; they directly improve the next team formation and collaboration.

Experimental Validation: Collaboration Evolution Delivers Real Gains

Experiments cover GAIA, LoCoBench, LOCA-Bench, and BeyondSWE across six metrics, with benchmarks interleaved into a continuous task stream to test continuous evolution in heterogeneous environments.

Under the mixed-task setting, EvoTeam achieves the highest scores on all six metrics, with an average score of 59.58 , improving 4.03 points over Meta-Team's 55.55.

Ablation studies break down the sources of gain. On GAIA, the full system scores 77.00; removing interaction rules drops to 67.00; further removing reusable teams drops to 61.00. On LOCA-Bench, corresponding scores are 67.08, 58.33, and 55.83, confirming that both team organization and interaction rules contribute practically.

Cross-model transfer is tested using GPT-5.6-Luna for evolution and DeepSeek-V4.1-Flash for evaluation. Across six comparisons on GAIA and two LoCoBench tasks, the gap between cross-model and same-model evolution never exceeds 4.00 points, demonstrating cross-model transfer potential of collaboration experience.

EvoTeam: Multi-Agent Collaboration via Evolving Teams
https://github.com/DEEP-JLU/EvoTeam
https://arxiv.org/pdf/2610.39813v2