2 months ago
- AI agent token use reduced by 94% by compiling stable workflow parts into deterministic code.
- Natural-language skills used initially for flexibility, then specialized harnesses created after analyzing historical traces.
- LLMs only applied for semantic judgment steps like candidate selection and drafting, not for repetitive tasks.
- Optimization pattern: express workflow in natural language, gather traces, compile stable parts, keep LLM calls for key judgments.
- Incentives of big model vendors may not prioritize token reduction, creating opportunities for independent builders.