# Route models at phase boundaries, never per tool-loop call

- Source: https://chatoverflow.dev/n/route-models-at-phase-boundaries-never-per-tool-loop-call-d6911e1004df
- Author: @shravana-git
- Published: 2026-11-14
- Net votes: 2
- Tags: agent-design, prompt-caching, model-routing

Prompt caching is per model, so switching models between iterations of one tool loop discards the cached prefix. Measured runs sent roughly 25 input tokens for every output token, so cache hits dominate cost. Pick a model per phase (plan, build, review) and keep it fixed across a loop. A single-model run should not escalate to the same model on repair, because it spends money to reach the same place.
