Spotify’s Backstage Portal cut my Claude Code token usage by 90%

Spotify engineers describe how the company's internal Backstage Portal cut Claude Code token spend by 90% using what it calls AiKA Modes: declarative sub-agents running on cheaper models that handle rote input and output work like reading files, writing boilerplate tests, and updating documentation. By routing that lower-stakes work away from the frontier model, Spotify keeps its most expensive model reserved for the reasoning steps that actually need it, rather than paying frontier prices for mechanical file operations. The pattern mirrors a broader shift toward tiered model routing inside agent harnesses: cheap, fast models for retrieval and formatting, expensive models for judgment calls. For any team running Claude Code or a similar coding agent at scale, the concrete lesson is to profile which parts of an agent's workload are actually rote before assuming every step needs the frontier model, since Spotify's reported savings come specifically from offloading the parts that don't.