Pattern: The Harness Now Moves Coding-Agent Results More Than the Model

Digital Applied·medium signal

Across June 2026 benchmark analyses, the dominant variable is scaffolding, not raw model choice — the same model family scores 51.9% vs 69.2% depending on harness, and one lab quietly stopped reporting SWE-bench entirely. The practical implication: investment in tool design, context engineering, and verification loops yields larger gains than swapping in a 'better' model. Teams optimizing agent performance should profile and iterate on their harness before chasing model upgrades.