Most AI coding agent work is I/O, not reasoning. Reading files, generating tests, updating docs—thousands of tokens burned on frontier models overqualified for the task. By 2028, AI coding costs are expected to exceed developer salaries, with engineering leaders spending $200–$500+ per developer monthly on tokens.

Portal by Spotify offers a solution with two modes: bulk-reader for file analysis and code-writer for boilerplate generation. Both use Gemini 2.5 Flash but can accept any configured model. Modes are declared agents running on ephemeral runtimes—no infrastructure management required.

The current implementation uses Claude Code plugins (shunt) with Pre Tool Use hooks that check file sizes and route large reads to bulk-reader automatically. This eliminates manual routing rules and ensures consistent delegation across projects.

The key insight: route grunt work to cheaper models and reserve expensive ones for complex reasoning tasks.