An agent can produce code quickly, but at the accuracy level of a co-op unless a significant amount of time is spent to "prepare" the agent. Trouble is, you have to know how to do the job you're asking the agent to do to be able to effectively steer it to an acceptable solution.

With a co-op (or junior engineer), I still need to do the equivalent of the above to train them on the tasks/concepts they need to perform their work, I still need to have a firm grasp of what it is that I'm asking them to do to be able to guide them to the correct solution.

The difference is that the co-op/junior retains that knowledge and is able to extrapolate that to different situations as well as come up with innovative solutions based on that initial knowledge. One of the biggest risks to tech work from LLMs (aside from the general problems with LLMs) is it encourages managers to not hire juniors/co-ops, which means that there will be no new generation of senior developers coming up to replace us when we retire (lol, as if).