According to CoinMeta, as reported by AI, AI has begun to automatically modify the agent's harness. scale has launched harnessopt-bench specifically to test whether AI can self-optimize. During the testing process, AI will be given an intentionally under-optimized harness, and the code will be continuously improved based on test scores and failure cases. After the modifications, scale will use unseen test questions to verify whether the improved agent has truly become stronger. This test utilized five models: claude, opus, gpt-5.6, SOL, and kimi, and 111 tests were conducted across four types of tasks. Among them, opus performed the best, winning first place in three out of the four categories. Overall, the model responsible for modifying harness is even more important than the coding tools used.