Harness Engineering vs Platform Engineering: A Mental Model for How Both Win

Two posts in early 2026 reframed how the industry thinks about agent quality. LangChain’s Deep Agents team published harness engineering numbers that moved their coding agent from outside the Top 30 to rank 5 on Terminal-Bench 2.0, a 13.7-point gain achieved by holding GPT-5.2-Codex fixed and changing only system prompts, tools, and middleware. A few weeks later, Cursor published a long read on how its harness team manages context, classifies errors, runs continuous A/B tests, and tunes per model family. Then came the Cursor SDK, exposing the harness as a programmable surface to anyone.