On March 6, 2026, TheGP hosted Ryan Lopopolo @ OpenAI to share his early experiments with harness engineering and agent-first codebases. Below you’ll find the summary notes, recording, and transcript from that conversation.

Please keep confidential — OpenAI has asked us not to share beyond TheGP Portfolio.

https://customer-0egndxlpnhogqmgf.cloudflarestream.com/a625a056ed1e37c36b09541962f00a00/iframe?poster=https%3A%2F%2Fcustomer-0egndxlpnhogqmgf.cloudflarestream.com%2Fa625a056ed1e37c36b09541962f00a00%2Fthumbnails%2Fthumbnail.jpg

Summary

Ryan’s core argument is that if you treat agents as real teammates and design your codebases for them, you can get an enormous amount of leverage—but only if you’re willing to be extremely opinionated about architecture, process, and documentation.He keeps coming back to a few big themes:

1. What “harness engineers” actually are

He thinks the future software engineer is less a line-by-line coder and more a systems designer and product thinker. The job is to:

• Be very product‑minded: start from user journeys and “definition of done,” not from functions and files.
• Think in systems: watch where agents get stuck or make mistakes and then redesign the system—docs, lints, patterns, workflows—so those mistakes become impossible or at least unlikely.
• Bring diverse expertise: the team should cover backend, infra, frontend, etc., so the harness (tests, patterns, guardrails) is strong across the whole stack. If you only have backend folks, you get, as he put it, grotesque React hooks and weird front‑end patterns.

2. The day‑to‑day rhythm: delegate, then disappear

He aims to behave with agents the way a good lead behaves with humans:

• Queue work up front, then walk away and wait for a PR instead of “shoulder‑surfing” the agent.
• Use PRs plus trust signals (screenshots, logs, QA evidence) rather than micromanaging every step.
• Spend his scarce synchronous time only on: (a) truly ambiguous, net‑new product work where agents have no prior patterns, and (b) deep refactors / architecture surgery.

Because his own time is not “consumed” in the act of coding, he’s happy to throw away 50–90% of agent output; there’s no sunk‑cost fallacy.

3. Codebases built for agents, not humans

He deliberately shapes the repo to be maximally legible and enforceable for agents:

• The code is extremely uniform: same patterns, same lint rules, 100% test coverage, strict limits on file size, etc., across hundreds of packages.