Four agents did not clear Astro's backlog because there were four.
They cleared it because none was allowed to be the hero.
In teams I have led, the most fragile workflow is the one where the same person interprets the problem, implements the fix, validates it, and declares it complete. Giving that shape to one AI agent does not remove the weakness — it accelerates it.
Astro separated reproduction, diagnosis, verification, and implementation, with each stage leaving evidence for the next and the original reporter testing the preview. No agent owned the issue end to end; labels and generated evidence carried the state between stages.
The useful design principle is not "add more agents." It is "do not let one actor manufacture and approve its own reality."
Agent count is an implementation detail. Separation of duties is the control.