Props to @_can1357 for this deep dive into the harness problem. The entire AI coding debate focuses on which model is best—GPT vs Claude vs Gemini—when the real bottleneck is something far more mundane: the edit tool. Most coding agents use str_replace (exact text matching) or apply_patch (diff-based). Both force models to reproduce content perfectly, including whitespace. Result? Grok 4's patch failure rate hit 50.7%. These aren't bad models—they just can't speak the language the harness demands. The solution? Hashline: tag each line with a 2-3 char content hash. Models reference the hash instead of reproducing exact text. One afternoon of benchmarking across 15 models: Grok Code Fast: 6.7% → 68.3% (10x improvement) Gemini: 8% success rate Output tokens: -61% (fewer retry loops) Zero training compute. Just better tooling. The gap between "cool demo" and "reliable tool" isn't model magic—it's careful engineering at the tool boundary. Yet vendors are blocking open-source harnesses instead of learning from them. The harness problem will be solved. Question is: by one company, in private, for one model? Or by a community, in the open, for all of them? Full breakdown: blog.can.ac/2026/02/12/the-h…