Descriptions:
Ryan Lopopolo, a product engineer on OpenAI’s Frontier team, joins the Latent Space podcast to define and demonstrate ‘harness engineering’—the practice of building production AI systems in which the human developer writes zero lines of code. Over five months, Lopopolo’s team built an internal tool with a codebase exceeding one million lines without personally authoring any of it, relying entirely on Codex CLI running at an estimated one billion tokens per day. He argues the models and harnesses have become isomorphic to a senior engineer in capability, making the harness itself—not the model—the primary engineering surface worth investing in.
The episode covers the mental model shifts required to work this way: abandoning instincts toward human-legible debugging tooling, thinking at the systems level about where agents make mistakes, and building confidence in automated loops rather than keeping humans in the critical path unnecessarily. A central technical concept is the ‘ghost library’ spec format, where a Codex agent reads an existing proprietary repository, writes a natural-language spec, a second agent implements the spec in a clean environment, and a third reviews the implementation against the original—looping until the spec can faithfully reconstruct the system from scratch.
Lopopolo also previews Symphony, OpenAI Frontier’s upcoming enterprise agent deployment product, and discusses the broader arc from early Codex mini experiments to today’s long-horizon autonomous coding agents. The conversation is among the most technically grounded first-person accounts of AI-native software development at production scale published to date.
📺 Source: Latent Space · Published April 07, 2026
🏷️ Format: Interview
