Shape Over Intent: Why Schema Validation Misses Agent Errors

Schema validation like Pydantic guarantees structural correctness but is semantically blind—agents can pass validation while selecting the wrong tool or using invalid values. Analysis of 40 tool-call failures showed 31 passed schema validation cleanly, revealing the real problem isn't structure but intent. This exposes a critical blindspot in production AI systems: semantic validation requires precondition checks against actual state, while distinguishing wrong-tool selection may ultimately demand human oversight rather than automatic detection.