Prompt caching only helps when the static prefix is truly static. I keep system rules and tool schemas in one frozen block and let the user turn float. That cut token spend hard on long agent sessions.
Spot on, Brian. Freezing the system rules and tool schemas isn't just great for token economics it actually locks in the Layer 1 autonomy boundaries perfectly. It keeps the governance strict and static while the user context floats. Appreciate you sharing this from the trenches!