I got lucky I think, in which the workflow of using agents is quite fun and addictive for me. But to audit anything you need a basic level of understanding of the underlying architecture, because we aren't there yet to just leave the agent working unattended and have plausible results, so any audit requires careful human involvement.

It's just that... Take this tiny project of mine I did for the fun of it https://gitlab.com/gabriel.chamon/thoracic-atlas-viewer. Basically went to an online interactive thoracic atlas and saw that the files were 404ing for download, but the visualization worked, so I gathered evidence and asked the agent to rebuild the archive for me. I still need to hack together an offline viewer, but it's amazing how much you can do just by having a general idea of what you need to do.