Do you give AI agent the specs and have it start building unattended?

Question

I use Cursor and I usually put it in plan mode and feed it the specs so I can review the plan and make sure the implementation will be done the way I want it to or to make sure the plan "feels" right.When I hear or see that people feed the specs, let agents handle things and then do some quick tests to make sure things work as expected, that sounds way too risky. I still have bad experiences where the agents made bad assumptions and included things I didn't need or ask for or had to stop the agent during implementation because it was going down a path that was needlessly complicated.

bradleyy · Accepted Answer

I'm using "harness engineering" to do this: smaller tasks, well defined stop conditions, runs in a VM with --yolo-mode. I've worked up to this, and ended up rolling my own thing because nothing I found did what I wanted:
* agent fan-out * VM containment * full harnesses with test/exit conditions * runs fully unattended
I expend a lot more energy on plans, though. I've got a full "Karpathy knowledgebase" that's effectively my jumping off point, so some of my projects have literally dozens or hundreds of docs they can source to build plans.