Principles / The Validation Gate
§ Principle 17 of 23
The Validation Gate
No agent output should reach production without structured evaluation against explicit criteria.
The Validation Gate is the Separation Principle applied to quality. Just as planning and memory should be separated from the model, evaluation should be separated from the agent being evaluated. An agent evaluating its own work is like a student grading their own exam — there's an inherent conflict of interest. The Validation Gate inserts an independent quality check between agent output and production deployment, using explicit criteria (not subjective assessment) to determine whether the output meets standards.
Why it matters
Without a Validation Gate, Execution Hallucination can reach production. The agent says "done," nobody independently verifies, and the gap between claimed and actual completion isn't discovered until a user hits a bug or a system fails. The Validation Gate catches these failures before they reach production — not by catching every error, but by ensuring that at least one independent assessment occurs before deployment.
In practice
For every agent workflow that produces production-bound output, define: what criteria must be met? Then implement automated evaluation against those criteria. The evaluation should be independent of the agent — ideally using a different model, different prompt, or deterministic checks.