Never trust an LLM's output directly. Here's the validation layer I put on every agent.

Short summary

Don't blindly trust LLM outputs—structured mode doesn't guarantee semantic correctness. Implement a three-stage validation pattern (parse, validate, classify) using TypeScript and Zod to force proper error handling and distinguish transient failures from permanent ones.

• •JSON validity ≠ semantic correctness: valid JSON can contain made-up UUIDs, invalid floats, or hallucinated structure
• •Build parse-validate-classify stages between raw model output and production code
• •Use discriminated unions to force callers to handle both success and specific failure modes

Generated with AI, which can make mistakes.