Speed without quality loss only holds when review load stays flat. I cap every agent PR at one concern and bounce anything that touches two modules at once.
That’s the missing link for SDD. Tools like GitHub Spec Kit get the agent to break things down into atomic tasks, but human review is where the one concern rule is actually enforced. If we let the agent bundle modules, the spec becomes too high-level to catch edge cases. Capping the PR scope keeps the verification tight.