Agent orchestration, evaluation, and efficient reasoning

CEPO reports prompt-optimization gains across four optimizers and 11 benchmarks; RoboArchitect explores persistent memory and evolved strategies around frozen robot policies, while HALO evaluates proactive-agent restraint across 50 environments. These complement earlier recursive self-play and efficient-reasoning proposals; effect sizes, benchmark scope, and real-world robustness remain key checks.