You get a feel for how much direction they need after working for a while and tooling and accessible documentation is really important for quality.
Then you give them a task and review the results. In (backend/systems) programming it's pretty binary whether a solution works or not, it's not a matter of taste but something you can just validate with hard data.
I've done so many tiny/small/medium sized utilities for myself in the last year it's crazy[0]. A good bunch of them are 95-100% vibecoded, meaning I was just the "project manager" instructing what features I want and letting the agent(s) make it work.
I think I have a pretty good feel for the main agentic systems and what they can do in the context of what I do so I know what to tell them and how - each has its own distinct way of working and using the wrong one for the wrong job is either stupid, frustrating or just a waste of time.