Agents2h ago
RSEA Agents Evolve Via Held-Out Selection
The RSEA framework uses a three-layer natural-language state to evolve strategies, skills, and playbooks without weight updates. It commits new candidates only after passing a strict keep-better gate on disjoint held-out data. This approach prevents the common pitfall of overfitting to a single benchmark. Practitioners gain a more reliable method for autonomous agent improvement.