SecondBrain
Ask the Brain
Index/Entityupdated Sat May 09 2026 08:00:00 GMT+0800 (Philippine Standard Time)

Ryan Lopopolo

personopenaiharness-engineeringlopopolo

Ryan Lopopolo

Member of technical staff at OpenAI. Self-described token billionaire (>$1,000/day in output tokens). For the last nine months, has been building software exclusively with agents — has banned his team from touching their editors. Author of OpenAI's "Harness Engineering" article (https://openai.com/index/harness-engineering/) and the corresponding talk.

Why he matters in this wiki

Lopopolo is the OpenAI counterpart to Boris Cherny — same lived experience (full-send agentic engineering), different harness (Codex instead of Claude Code), different framing. His talk is the most operationalized recipe in this wiki for what it actually looks like to drive a team of agents day-to-day.

He coined or named several terms now used elsewhere in the vault:

  • Code Is Free — implementation is no longer the scarce resource
  • LLM as Fuzzy Compiler — code as a build artifact, context as constraints
  • Garbage Collection Day — Friday ritual: turn every PR slop you saw that week into a permanent guardrail
  • "Every time I have to type 'continue' to the agent is a failure of the harness."

Practices on his team

  • No editors. Codex is the entry point; everything (app launch, observability, Chrome DevTools) is wired in as a skill so agents can drive it.
  • 5–10 skills, deepened over time — not hundreds. He hides infrastructure churn behind them so humans don't track it.
  • 750 packages in a PNPM workspace. Heavy use of package privacy + dependency-edge lints to give agents structural hooks even without microservices.
  • Reviewer agents triggered on every push, prompted with persona docs (front-end architect, scalability, reliability) — replaces synchronous human code review.
  • Tests about source code in addition to tests of code behavior — e.g. "files must be ≤ 350 lines" — to keep the codebase context-efficient.
  • Workflow: kicks off a task, tethers laptop to phone, buckles laptop into the back seat of his car, lets it cook during the 30-min commute home.

Cross-link to Patrick Debois

Same conference (AI Engineer London 2026), companion talks. Patrick describes the Context Development Lifecycle — the discipline around context. Ryan describes Harness (LLM Agents) engineering — the discipline around the agent runtime. Both arrive at the same primitives (LLM as Judge as the testing mechanism, packaged skills/lints as the distribution unit, logs/PR comments as the feedback signal).

Sources