Agents3h ago
RSEA Agents Evolve Via Held-Out Selection
A three-layer natural-language state allows RSEA agents to rewrite strategies, skills, and playbooks without updating model weights. The system employs a strict keep-better gate to prevent regression on disjoint held-out data splits. This approach stops the common trend of overfitting to single benchmarks. Practitioners gain a more reliable method for iterative agent optimization.