A Better Model Won’t Fix Your AI Agent. A Better Harness Will
中文摘要
优化测试框架比单纯升级模型更能提升AI智能体性能。GPT-5.2-Codex仅通过框架改进,就在Terminal-Bench上大幅提升排名。
English Summary
Better harness engineering, not just better models, improves AI agents, as evidenced by GPT-5.2-Codex's significant rank jump on Terminal-Bench through harness changes alone.
Original Excerpt
A frozen GPT-5.2-Codex jumped 25 ranks on Terminal-Bench, +13.7 points, on harness changes alone. This is harness engineering. Continue reading on Level Up Coding »