Tiering coût/qualité (advisor, delegate, escalate)
Cost/quality tiering means you stop sending every call to the most expensive model and route each one to the model that is enough. Anthropic names three patterns for it in its « Building Effective AI Agents » cookbook. The advisor: the small model does the work and only consults the big one when it gets stuck, roughly once per task. The delegate (orchestrator-workers): the big model plans and hands the bulk of the execution to cheaper workers, becoming the manager who no longer types. The escalate: you only move up a tier when the task genuinely demands it. In all three cases the goal is the same, staying close to the big model's quality while paying only a fraction of it.
Strengths
- Three patterns named and documented by Anthropic, not a homegrown hack
- Real drop in the bill on mixed tasks (a lot of easy, a little hard) with no perceptible quality loss
- Usable today in Claude Code: an advisor sub-agent or delegated workers, with no homegrown orchestration layer
Limitations
- On a uniformly hard task, routing buys you nothing: you will pay for the big model either way
- Poorly calibrated (an advisor that escalates at every step), you pay the double call and never reach the saving
Best for
- Devs running agents and watching the bill climb on mostly trivial steps
- Tech leads who must decide which model goes on which step before scaling it up