TRACE-EVO: Contribution-Driven Role Reassignment and Knowledge Transfer in Long-Horizon Multi-Agent Systems

Large language model agents are increasingly organized as teams with specialized roles, persistent memories, and reusable skills. Yet most multi-agent systems keep role assignments fixed or optimize only the final answer, making it unclear when accumulated experience should trigger organizational change. We study long-horizon heterogeneous multi-agent systems where agents differ in base capability and acquire task-specific knowledge over repeated work. We propose TRACE-EVO, a framework that treats team operation as a closed-loop organizational control problem. TRACE-EVO estimates each agent's process-level contribution, builds role and agent profiles from verified outcomes, searches for capability-first role reassignments, and transfers only validated skills and memories through trial, verification, and rollback. To evaluate this setting, we construct TRACEBench, a suite of stateful enterprise-style tasks spanning code, data, incidents, evidence, product release, and operations, with growth phases followed by frozen transfer tests. Experiments with heterogeneous LLM teams show that static role layouts and shared-memory baselines often fail under role mismatch and task drift, while TRACE-EVO improves downstream team performance and reduces catastrophic failures without changing any agent's underlying model. Our results suggest that long-horizon agent teams should be evaluated not only by final answers, but also by whether they can identify contribution, preserve useful organizational knowledge, and reassign work as capabilities evolve.