Stanford Researchers Introduce TRACE: A Capability-Targeted Agentic Training System That Turns Recurrent Agent Failures Into Synthetic RL Environment

TRACE is a training system that improves AI agents by identifying specific skills they repeatedly fail to perform and then training targeted expertise for each missing capability. The system works by analyzing failed agent attempts to find common skill gaps, generating synthetic environments that isolate each gap, training specialized adapters for those capabilities, and routing tokens across the appropriate expert during operation. This approach matters because mainstream training methods either provide sparse feedback that does not indicate which skill was missing or generate broad synthetic data that wastes resources training skills the model already has. Testing showed TRACE achieved substantial improvements over baseline methods on customer service and software engineering tasks while using fewer training examples than competing approaches.