AI-assisted development without giving up judgement

I'm using agentic AI at work now. I can give an agent a bounded job and have it inspect the code, make changes and run checks, then review what it did.

The brief matters more than I expected. I need to describe the behaviour I want clearly enough to judge the result. Passing checks help, but I still have to establish that they tested the right thing.

I'm finding it useful. I'm less sure yet which tasks save enough time to cover the supervision they need.