Developer cuts Claude Code token use 90% with AI agent modes

write-up · for knighthk-irc in #systemcrafters · 2026-09-05 03:50 UTC

Portal’s AI agent modes cut this developer’s Claude Code token use by 90% by routing routine file-reading and boilerplate-writing to a cheap model.

AI coding agents spend most tokens on grunt I/O like reading files or generating repetitive test code, not actual reasoning — burning frontier-model tokens on that is wasteful.
Portal’s "modes" are declarative agents on an ephemeral runtime; you define instructions, pick a model, and attach tools — no infra or API keys needed, callable from CLI or API, public or private..
A "bulk-reader" mode (Gemini 2.5 Flash) handles one-off file-reading/questions with terse, bullet-only answers, skipping all preamble..
A "code-writer" mode generates predictable code (tests, config, stubs) matching existing patterns exactly, emitting only raw code with no markdown fences or prose..
Routing evolved from advisory CLAUDE.md rules (which Claude could ignore) to a Claude Code plugin called "shunt" that enforces the cheap-model delegation instead of relying on self-routing..

Commenters are largely skeptical of the technique, framing it as a standard multi-model delegation pattern rather than a breakthrough, with one noting it merely shifts token costs to a cheaper service and another questioning whether it actually saves tokens once the big model must plan, review, and distribute work. The most notable critique is the absence of accuracy or performance benchmarks, with one user calling for evaluation on a real coding benchmark like DeepSWE bench instead of just token savings. Real disagreement emerges over the value of using a "dumb" model as a smart grep—some see it as a practical cost-cutter, while others argue the whole point of frontier models is their intelligence on hard code, not on reading files. A separate thread of frustration targets the site's hijacked scrolling and the "AI slop" nature of the article itself, though these are side complaints rather than central to the technical debate.

See also