Comment by pym4n

You missed a key point here: context bias. It's always better to have conpletely separate agents per task if you want "fresh eyes" on any given task.