The real test of agent orchestration is can you have a very intelligent model (i.e. Opus-4.6) hand-hold a very stupid model (i.e. Llama-3) into producing correct results. (Through perfect task decomposition, hyper-specific prompting, and relentless verification/retry loops). You'd never do this in production, obviously use a much smarter 'dumb' model where cost or tuning is a factor. But IMO this is an honest stress test; if your orchestration only works when the underlying model is smart…