The OpenAI frontier team shipped 1 million lines of production code. Zero written by humans.
What's interesting isn't the scale—it's the mechanism. They encode taste via testing. Every qualitative judgment the design team makes gets compiled into CI. Typography preferences, component rules, doc cross-linking requirements. If the output violates any of them, the build breaks.
Ryan Lopopolo frames it simply: "Think of it as you just hired a cluster of interns named Codex. Your job is to manage them."
That reframe hit me. The 5-layer harness is the management layer. agents.md defines the operating loop. Persona-based CI reviewers grade every diff. The model gets better every release. The harness gets better every failure. Both compound. One of them you control.
I keep turning this over. The product isn't the code anymore. The product is the system that manages the entity writing the code.
What does that make you, really?
Top comments (0)