EvoTeam

Roles and collaboration strategies in multi-agent teams are usually fixed by developers. EvoTeam makes constrained changes to a versioned Strategy based on cross-task execution evidence, and a candidate must pass a validation gate before it is promoted. The main scenario is generating and checking complex project plans.

The initial Strategy v0 is Planner → Executor → Critic, with a developer-defined role pool. The system runs the current strategy; repeated failures, performance drift, cost anomalies or low-contribution evidence above a threshold trigger offline evolution; once stable, the strategy enters STABLE and the system stops exploring but keeps monitoring.

Online:  Task → Current Strategy → Orchestrator → Team / AgentInstance
         → Run → Evaluator → SealedRun → Experience → StrategyMonitor

Offline: Trigger → EvolutionManager → Attribution → Mutation → Candidate
         → Validator → Improvement Attribution → Validation Gate
         → Promote / Reject → later tasks use the validated Strategy

• Integration of the real ReActAgent from openJiuwen 0.1.17.post1.
• init / run / observe / evolve entry points, SQLite traces and snapshots.
• Repeated-failure monitoring; two candidate types (prompt changes and an added Verifier) and multi-candidate comparison.
• Regression runs with DeepSeek covering complex tasks, bounded rework and real gate rejections of multiple candidates.

• Drift, cost and contribution monitoring.
• Tool / skill policies and more general structural mutations.
• Experiment thresholds still need baseline calibration.