Spotify's Portal offloads Claude Code grunt work to cheap models, cutting tokens 90%

A Spotify engineer argues that most of what an AI coding agent does is I/O, not reasoning — reading files to answer a single question, scaffolding tests that mirror existing ones, updating docs — and that pointing a frontier model at all of it wastes money. With token spend projected to eventually rival developer salaries and some teams already burning $200–$2,000 per developer per month, the pitch is to route predictable, low-reasoning tasks to a cheaper worker model and reserve the expensive one for problems that need it.

The implementation uses Portal by Spotify’s AiKA Modes, declarative agents that run on ephemeral serverless runtimes with no infrastructure or API keys to manage. Two modes do the work: a bulk-reader that ingests large files and returns terse structured summaries, and a code-writer that generates boilerplate matched to a required reference file. A Claude Code plugin called shunt enforces the routing through three layers — PreToolUse hooks that block reads of files above a configurable line threshold (default 350) and redirect them, bash scripts that wrap the Portal CLI and keep the file contents out of Claude’s context entirely, and skill files that tell Claude how to invoke the scripts. Because the heavy content never enters Claude’s context, benchmarks against a Java monorepo showed roughly 90% token savings on reads, and generated code goes straight to disk unseen.

The approach has clear limits. Editing can’t be delegated because the worker’s summaries lack reliable line numbers, so targeted reads still pass through. Reasoning can’t be delegated either — the cheap model missed a subtle thread-safety bug that Claude caught immediately — so debugging, architecture, and safety-critical code are explicitly excluded. Each delegation is a 10–30 second network round trip capped at 30 seconds, making it worthwhile for large reads but counterproductive for small ones, which is why the line threshold exists in the first place.