I find this interesting:

  …no model performed better with an Agent, a couple performed worse, and time/tokens/costs were consistently much higher with the agent in the loop, for some reason.