Harness engineering

"Agent = Model + Harness. If you're not the model, you're the harness." Viv Trivedy's equation, quoted by Addy Osmani (Google) on 19 April 2026, is the working definition: the harness is everything added around the model to improve what it generates. Osmani's supporting evidence comes from Terminal Bench 2.0, where Claude Opus 4.6 running inside Claude Code scores well below the same model in a custom harness, and where Trivedy's team moved a coding agent from Top 30 to Top 5 by changing only the harness.

OpenAI reported in February 2026 that three engineers (later seven) produced roughly a million lines of internal beta product in five months with zero lines written by hand, across about 1,500 Codex pull requests. Their conclusion was that the bottleneck on agent performance usually sits in environment design rather than model intelligence, and their three pillars were Context Engineering, mechanically enforced architectural constraints, and recurring garbage collection tasks. Birgitta Böckeler split the same object into guides (feedforward: conventions, specs, skills, MCP servers) and sensors (feedback: linters, tests, static analysis), each either computational or inferential. Rémi Jacquet of Cast Software dated the pivot: prompt engineering in 2024, Context Engineering in 2025, harness engineering in 2026.

Dorra Bartaguiz supplies the counter-measurement. Code generated from a single prompt was announced at 100% coverage; she measured 33% overall and a mutation score of 13.82%. Böckeler's hardest category stays open: verifying that an application does what it should.