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

Ryan Lopopolo of OpenAI argues implementation is no longer the scarce resource and explains how to operationalize agent-first engineering teams via harness engineering.

• Lopopolo banned his team from touching editors for 9 months; all code must go through agents.
• He spends over 1 billion output tokens per day (~$1,000+/day), self-described as a token billionaire.
• GPT 5.2 was his inflection point: models became isomorphic to human engineers in code quality.
• His team enforces a 350-line file-size limit via a test — explicitly adapting the codebase to model context constraints.
• They run 750 packages in a PNPM workspace, intentionally over-architected so agents have scoped, consistent subtrees to navigate.
• Fridays are dedicated garbage collection days: one day per week to categorically eliminate slop patterns observed in PRs that week.
• Token usage splits roughly one-third each across planning/ticket curation, implementation, and CI-based review agents.
• Code is treated as a disposable build artifact compiled from specs; swapping models is like changing a compiler backend.

2026-04-17 · Watch on YouTube