I Improved 15 LLMs at Coding in One Afternoon. Only the Harness Changed.
Summary
The article argues that the harness—the interface between the user and the model—is the real bottleneck in LLM-driven coding, not the models themselves. It surveys editing tools (patch-based, str_replace, and neural merge) and introduces Hashline, a tagging approach to anchor edits with content hashes, supported by benchmarks across 16 models. It advocates open, community-driven harness innovation over vendor-tuned solutions.