In my (limited) experience AI agents seem to check for errors by trial and error rather than logic.
Recently I was presented with a ‘solution’ simply because JS code had run a small number of times without error…
Similarly working with Swift it compiles therefore it works OK.
Recently SWE-2 failed to spot a logic problem that was obvious to (even) me at first sight.
Am I using the wrong tool or should I accept that eternal vigilance is the price of error free code?
Happy to provide detail if anyone is interested but I suspect it is a wider issue.
