Extreme Harness Engineering for Token Billionaires: 1M LOC, 1B toks/day, 0% human code, 0% human review — Ryan Lopopolo, OpenAI Frontier & Symphony

The conversation centers on the emerging discipline of "Harness Engineering," where developers use AI agents to automate the entire software development lifecycle (SDLC). Ryan Lopopolo, from OpenAI’s Frontier team, details his experience building a million-line-of-code Electron application with zero manual code, relying instead on Codex to manage everything from product features and testing to infrastructure and documentation. By treating code as disposable and prioritizing agent-legible architecture, the team achieved a 10x increase in productivity. Key strategies include enforcing strict build-time limits, using markdown-based specifications to guide agent behavior, and implementing self-correcting loops where agents review their own PRs. This methodology shifts the human role from writing code to systems thinking—designing the constraints, observability, and feedback loops that allow autonomous agents to operate reliably at scale.