Agent Evolution

The paper interprets wall-clock bottlenecks as systems-level orchestration problems rather than only sample-efficiency problems. Agent evolution methods use LLM reflection over execution traces and avoid labeled trajectories and gradient updates. LLM-based agent evolution improves behavior by refining non-parametric artifacts rather than changing model weights. Synchronized agent evolution creates inefficiencies from sequential stage dependencies and workload imbalance within stages. The practical bottleneck in existing agent evolution approaches is wall-clock time.