It s been a couple weeks now since I ve begun doing it the other way around. Instead of inspecting the function itself, I begin by reading the tests that the agent generated for it. These tests reflect the agent s understanding of what the code is supposed to do, which is not necessarily what I asked for.
Twice now I found an oversight in ten seconds that otherwise would take an hour of debugging. The code was fine both times. The tests pointed the issue out.
