David's approach: use subagents when the work is actually orthogonal

David often asks Codex or Claude to launch multiple high-effort subagents for broad investigations: one can read logs, one can inspect frontend state, one can inspect database rows, one can audit prompts, and one can compare deploy history.

but David also draws a boundary. if implementation is tightly coupled, one main agent should keep the whole model in its head instead of scattering changes across workers.

the rule is not "always spawn agents." the rule is: use parallelism when the questions are independent; preserve coherence when the system design is entangled.