Spotify's Portal AiKA Modes routes repetitive I/O tasks like multi-file reads and boilerplate code generation from Claude Code to cheaper models (e.g., Gemini 2.5 Flash), cutting token consumption by roughly 90% in Java monorepo benchmarks.

A Claude Code plugin called "shunt" enforces this via PreToolUse hooks that block reads of files exceeding a line threshold and redirect them to Portal-hosted agent modes, while a second mode writes pattern-matched code directly to disk without entering Claude's context. The system cannot delegate editing or reasoning tasks, and each delegation adds 10-30 seconds of network latency, making it net-negative for small reads below the threshold.