I've used agents quite a bit and I agree. The current baseline workflow is something like agent output -> human review -> agent refinement -> human review -> a

I've used agents quite a bit and I agree. The current baseline workflow is something like agent output -> human review -> agent refinement -> human review -> agent refinement -> ... But agents are capable of making meaningful improvements to their own output. I'm hoping dynamic workflows move towards something like: agent output -> agent review -> agent refinement -> (cycle to fixed point) -> final human review