I’ve noticed that custom subagents in CC often perform noticeably worse than the main agent, even when told to use Opus and despite extreme prompt tuning. This seems to concur with the “keep it flat” logic here. But why should this be the case?

(blogpost author here) I've noticed this too. My top guess for any such thing would be that this type of sub-agent routing is outside the training distribution. Its possible that this gets better overnight with a model update. The second reason is that sub-agents make it very hard to debug - was the issue with the router prompt or the agent prompt? Flat tools and loop make this a non-issue without loss of any real capability.