Dimitri Mazmanov’s core observation: most of what an AI coding agent does isn’t thinking — it’s I/O. Reading five files to answer a question about one method, generating tests that mirror the twenty tests next to them. That grunt work burns tokens on a frontier model that is wildly overqualified for it.
The cost pressure is real. Gartner predicts AI coding costs will pass the average developer’s salary by 2028, and a quarter of engineering leaders already spend $200–500 per developer per month on tokens, some past $2,000.
His fix is model routing with two cheap “modes” — declarative agents that run briefly on demand (think serverless, but for AI agents):
- A bulk-reader mode: given file paths and a question, returns tight structured bullets instead of dumping whole files into the expensive model’s context
- A code-writer mode: generates tests, config scaffolding, and stubs that match a reference file’s existing patterns exactly
- Both run a cheaper worker model (Gemini 2.5 Flash in the examples), and Portal handles infra, keys, and cleanup
The enforcement layer is a Claude Code plugin. Hooks block reads of files over a size threshold and redirect Claude to the cheap reader; scripts wrap the calls; skills tell Claude when and how to delegate. Measured against a Java monorepo, mean token savings on bulk reads were around 90%.
The honest limits matter as much as the savings: you can’t delegate edits (summaries don’t carry reliable line numbers), you can’t delegate reasoning (the cheap worker missed a subtle thread-safety bug the frontier model caught in seconds), and each delegation adds 10–30 seconds of latency, so small reads aren’t worth it.
The bigger idea is that routing decisions and worker behavior are now separate, reusable configuration — “when to delegate” lives in the plugin, “how to respond” lives in the mode. That turns token-cost management from a systems-engineering problem into a config problem any team can adopt.