Harness engineering

Three engineers at OpenAI, later seven, produced roughly a million lines of code in five months with zero lines written by hand; Codex agents opened, evaluated and merged around 1,500 pull requests, an average of 3.5 per engineer per day. The methodology behind that number, harness engineering, treats the environment around the model as the engineering object: repository structure, CI configuration, a strict dependency-layer model, custom linters whose error messages inject correction instructions straight into the agent's context, and recurring garbage-collection tasks that open auto-merged refactoring PRs.

Viv Trivedy's equation, quoted by Addy Osmani on 19 April 2026, gives the field its slogan: "Agent = Model + Harness. If you're not the model, you're the harness." Osmani cites Terminal Bench 2.0, where Claude Opus 4.6 scores lower inside Claude Code than the same model in a custom harness, and Trivedy's team, which moved a coding agent from Top 30 to Top 5 by changing only the harness.

Birgitta Böckeler splits the practice in two: guides, the feedforward controls (conventions, specifications, skills, MCP servers), and sensors, the feedback controls (linters, tests, static analysis). Each can be computational or inferential. The behaviour harness stays unsolved in her account, because AI-generated test suites are not yet reliable enough to verify that an application does what it should.

On BFM Business on 5 May 2026, Rémi Jacquet of Cast Software dated the pivot: prompt engineering in 2024, Context Engineering in 2025, harness engineering in 2026. His warning ran against the optimism of the segment: the stronger the AI gets, the more the guard drops.