Harness Engineering: How to Build Software When Humans Steer, Agents Execute — Ryan Lopopolo, OpenAI

Ryan Lopopolo, a member of technical staff at OpenAI, shares his experience building software exclusively with AI agents for nine months. He introduces 'harness engineering' — the practice of structuring codebases, documentation, and processes so AI agents can do the full job of software engineering with minimal human intervention. Key ideas include: treating code as free and disposable, making human time and model context the scarce resources to optimize, writing agent-readable documentation and ADRs, using custom ESLint rules and reviewer agents in CI to enforce non-functional requirements, structuring repos into many small isolated packages for context efficiency, and dedicating time to systematically eliminate recurring agent failures. He advocates for a future where engineers act as staff-level orchestrators delegating to parallel agents, and describes practical techniques like just-in-time context injection, source-code-level tests, and persona-based review agents.