From an engineering perspective: Agents are great if you are building prototypes or are flushing out an area of your stack that is heavily standardized (i.e. the AI has had a lot of examples to learn from, e.g. simple authentication flows, simple database singletons, basic QA automated testing, etc); However, if you are building new stuff or stuff the AI has not trained on, it is absolutley, infuriatingly prone to output garbage.

edit: it's no secret either, spin up a new project and you'll be greeted with this don't trust me billboard right from the get-go: