Armin Ronacher's "The Final Bottleneck" argues the bottleneck in shipping code has shifted from writing code to reviewing it, now that agents have become much more capable. AI generates code faster than we can review it, so backlogs grow and the system breaks when input exceeds what we can process.
He also links to an interesting tweet by Thorsten Ball: the bottleneck is often someone copying between issue tracker and coding agent. If AI can write both the report and the fix, why should a human still be stuck clicking through GitHub or issue trackers just to make the change happen?
I’m noticing the human bottleneck at work as well. These days, I can copy text from ten different Jira issues into ten separate Copilot agents and spin up ten parallel Codex 5.3 coding sessions. The results are almost always excellent. But as long as humans remain accountable for the final outcome, we still have to review every single contribution ourselves. We’ve achieved something close to a 100× increase in writing speed, but nowhere near that level of improvement in shipping speed.
The only way we’ll truly move faster is by letting go of the idea that we must personally own and fully understand all of the code. Right now, that feels like a terrible idea. LLM agents still make large mistakes with complete confidence from time to time—mistakes that could be disastrous if pushed directly to production.
If you see a company that suddenly ships a 100 times faster, you should be very wary.