DigiNews

Summary

Can Bölük argues that the harness—the interface around LLMs—controls real-world coding performance more than the model itself. He compares patch-based, string-replace-based, and hashline editing approaches, introducing a hashline tagging idea to anchor edits and reduce token waste. Benchmarks across sixteen models show that hashline-based edits outperform traditional formats for many models, with notable gains for open models; the piece also critiques vendor lock-in and advocates open harnesses as the best path forward for interoperable AI tooling.