Technical articles, tutorials, and insights
Where does an agent's cost actually go? Four experiments cover the core optimization strategies: system-prompt token breakdown (including Prompt Caching mechanics), model routing (direct LLM vs full agent), parallel tool calls (measured 3.0x speedup), and tool result caching (0ms vs 100ms). Counter-intuitive findings: latency with only 2 samples is noise, model routing has hidden overhead, and measuring always comes before optimizing.