David's approach: Delegation with parallelism: spawn sub-agents only for orthogonal work to protect coherence

Across multiple sessions, David articulated a clear rule for when to use sub-agents:

if the completion of the execution of the plan end to end includes orthogonal independent sections then you can launch gpt 5.4 xhigh sub agents that go and independently work on their sections that you partition them — but if the work is all related then you must do it all yourself so that the agents don't up the coherence of the system

This was applied explicitly when deciding how to execute the external-link file resolution plan for Kerra lecture notes generation. The concern was that agents working concurrently on interdependent code paths would produce conflicting changes that broke the overall system even if each individual change was correct. David distinguished between "orthogonal" (safe to parallelise) and "related" (must be sequential) as the operative criterion — not just whether sections touched different files, but whether their outputs had dependencies on each other.