Agent Coherence, Harness Engineering, and Evaluation Reliability

Research is shifting from static model scores toward durable harnesses, shared memory, verifiable environments, confidence estimation, self-testing, and realistic interaction. ScienceIDE, Agora, HypoEvolve, ProgramDistill, XConf, and new RL-critic analyses reinforce the trend, while delayed failures, reward hacking, weak transfer, verifier limitations, and uncertain cost-performance tradeoffs remain central problems.