An issue with Codex and Claude Code is that users need more control over the particular configurations of subagents that orchestrator AIs use.
I want to decide whether to delegate research or writing or user testing, and to which models. Otherwise it is a router problem again.