I personally explicitly avoid parallel agents, since it creates too much cognitive debt, and sometimes an agent may need steering towards an architectally sane solution mid-work.
My approach these days is to do one change at a time, until I can fully merge it with confidence.
With complex stuff I often have open the “thinking” output so I can stop/interrupt with guidance. Without that it’s often garbage output that I then have to work to fix, which is difficult to do while you’re also doing the same with a parallel process.