How do you guys trust AI enough to ship without checking what it built?
I'm running 5 to 8 agents in parallel, but I still need to check the final output before pushing to prod.
I know there are workflows and tools to automate this, but I can't seem to let go yet
Anyone else…
Commit
473237e
I ignore the code, but I test that the end-user experience is what I expect it to be. I managed teams of engineers, and this is no different, except when AI writes the code, AI can also fix the code, so not worried about syntax and function naming. No need for pair review.
(when…
YES Assaf, I am not talking about checking the code. I do not do that 90% of the time. I am talking specifically about the compulsion to check, even by asking what someone did and saying "ok." But even the checking you do is a bottleneck in a sense.
I was saying elsewhere that…
in my experience, you do have to do some manual checks, but typically that's around new features — you can teach the AI to do proper regression testing you just have to know where to start.
for example, yesterday I released three new features, didn't bother to manually test either. but today we're working on making web pages printable, and I have to teach the AI what it can hide in print mode, how to do visual regression tests, etc. so I would spent likely an hour today on that one feature.
sometimes I only have to look at very small parts of the code, eg when releasing API changes, I may only need to look at few lines of API schem, and trust the AI can write the code, the test suite, the docs, the changelog, etc.
commit 473237e