Agentic AI: Scaling Gains Meet Long-Horizon, Coordination, and Evaluation Limits

Evaluation is increasingly testing exploration, process validity, and deployment quality: AdaExploreBench exposes failures to seek evidence or detect rule drift, and OR-License finds a gap between strict feasibility and deployment-adjusted quality. Harness², CoworkEnv, and OpenVLHarness report gains in specific settings, but benchmark transfer, independent validation, and reliable long-horizon behavior remain unresolved.