P02: Model Routing
What You Do
Run the same prompt three ways: flagship model, small model, and a routed config. Compare turns, tokens, cost, and correctness.
Harness Mechanism
Remote-safe model selection before agent construction, with LLMRegistry as the larger-harness pattern.
Open First
Keep
A small routing policy: routine text and code-search prompts can use a cheaper model, while security, architecture, image, or multi-file tasks route up.
P09 returns to model routing as a full benchmark with RouterLLM, profiles, SwitchLLMTool, Laminar traces, and cost per solved task.