Verifying written code is still the bottleneck, even though the cost of writing it is negligible now. As the OpenAI team puts it: the people who used to be bricklayers, handwriting every piece of code, now spend their days as construction managers. This is a harder problem than it sounds, because a huge influx of written code brings with it a fraction of bugs and security holes baked in. At human pace, a 1% bad rate is noise you can catch. At 100x or 1000x the volume, that same 1% becomes a flood of problems moving faster than anyone can check. How the OpenAI team deals with this is interesting: instead of one checkpoint at the end, they filter at many points along the way: one agent writes the code, a separate specialist agent reviews it for its own domain, another watches how it behaves once it's live, and so on. Now that you can ship code at the speed of thought, I think the much harder part is fully trusting what you build.