An OpenAI engineer describes the weekly ritual that keeps a three-person team merging code unread

Lopopolo says he has spent nine months building software exclusively with agents, banning his own team from touching their editors, and reports usage north of a billion output tokens a day, over $1,000 worth by the moderator's estimate (19:00). Every Friday is "garbage collection day": his team converts every piece of PR-review slop it observed that week into a durable lint rule, a source-structure test, or a persona-scoped review agent, so the same class of mistake cannot recur (37:47 to 39:21). Other mechanisms: a hard 350-line file-length test enforced for the model's context efficiency rather than code quality (13:51), lint and test failure messages written as remediation prompts rather than bare errors (14:22), and a workspace of 750 packages isolated by business-logic domain specifically to give the agent filesystem-legible boundaries (34:26). He frames a model swap the way a compiler swaps optimization backends, LLVM to Cranelift, arguing the same acceptance rules should hold regardless of which model produced the patch (41:49 to 43:04). Auto-generated captions throughout (the ASR mangles his own name three separate ways), so everything above is paraphrased rather than quoted, and the token and package figures are his own unverified account of his own team's usage.

Why it matters: Lopopolo gives a first-party account of the weekly harness work required to keep humans out of the loop. The garbage-collection-day ritual is a process you can copy without copying any of OpenAI's own tooling.