Fail-Closed Eval Gate

A change to an agent or prompt is not shippable until an evaluation suite exists that can fail and block the release. If no eval artifact is present, or if the eval cannot fail, stop and build the gate first.

This is the CI equivalent for AI agents: the check that sits between code-complete and production.

When to run

• A pull request changes agent prompts, system instructions, tool schemas, or orchestration logic
• Someone says "ready to merge" or "ship this"
• A release candidate is proposed
• You're setting up CI/CD for the first time

If you can't name the eval suite and show its last failing run, the change is not gated.

The gate structure

Agent evaluation gates are tiered. Each tier blocks independently — a strong score on one tier does not carry a weak score on another.