Not Every Subagent Needs Your Best Model

We spent a week discovering that "always use the best model" is a bad default.

The trigger was mundane. One of us hit the top model's limit on a Tuesday, with the window resetting Saturday, while sitting at 9% of the overall quota for that window. That doesn't add up until you learn the strongest model runs against half the weekly limit rather than drawing from the same pool faster. It's a separate, smaller budget.

So leaving everything on the most capable model doesn't just cost more. It runs you out mid-week on work that never needed it.

Subagents are where this bites

Our Tailor Agent fans work out to subagents, and translation is the clearest case. Take a page's copy, translate it, hand it back. There's no architectural judgment in that. It's a well-specified transform with an obvious success condition.

We route those explicitly now. The instructions name Sonnet for translation subagents. The orchestrator, which decides what to change and why, stays on the stronger model.

The useful split is whether a task needs judgment about what to do, or just competence at doing it. Everything in the second category is a candidate for a smaller model, and there's more of it than you'd guess before you look.

Three things we'd tell ourselves a month ago

Route per subagent, not per session. Model choice is a property of the task, and a session is many tasks.
The expensive model is a separate budget, not a faster burn. Treat it that way and the mid-week wall stops being a surprise.
Latency counts too. A translation subagent on a smaller model finishes sooner. Across a fan-out of ten, that's the difference between waiting and not.

None of this is about frugality. A fan-out spends your best model on work a smaller one does identically, which is how you end up out of quota on a Tuesday.