I hate to break it to engineers feeling this for the first time, but this is what management and "leadership" has been dealing with forever in software. It's just you were a fleshy Claude/codex. And your goal was never to understand everything it was to build something of value that had the risk squeezed out of it through development processes.

Unfortunately the answer is that you better get comfortable with the feeling and know that when something goes wrong it's not the end of the world (for most of us, some of the AI labs it may actually be) and while it will be painful you fight through the incident, you do the post mortem and you don't let that class of error happen again.

There are some techniques, which you hated as an IC that you'll now have to embrace:

1. Have it write a doc and review it.

2. Have a strong point of view on critical parts of the architecture and one way doors. Review those mercilessly.

3. Spot check and taste test the code.

4. Be merciless about the small mistakes you see slipping in expand scope on them to find the class of problem it is.

5. Think about contracts and test coverage and failure modes at a higher level. What is the most effective way to find misalignment without reviewing every line.

Know that through this meta process you will squeeze out risk but that you won't eliminate it and at some point you feel like you squeezed out enough risk to ship.

And when something does go wrong take that moment to find a way to add another step to your risk reduction process that eliminates that class of errors not just that individual error.