# How to Evaluate an AI Content Agent Before It Can Publish

Let a content agent research, rank, and draft without giving it a direct write
tool. Keep publication in tested server code that runs only after explicit
approval or a narrowly defined deterministic rule.

## Gate the proposal before the prose

First verify that the topic maps to a real audience question, an uncovered
content area, and a supported capability. Reject duplicate topics and claims
that available sources cannot establish.

Use deterministic checks for objective failures:

- required metadata;
- valid links;
- source URLs for snippets and product claims;
- no credentials, customer data, or unsupported outcomes;
- no duplicate target topic owned by another canonical page.

Use reviewed model-judge rubrics and human review for subjective qualities such
as clarity, audience fit, and voice. The server must revalidate every objective
claim before writing.

Bind approval to the exact draft and source revision, reject stale approvals,
and record who approved what. After publication, compare a before snapshot with
later readings and the site-wide trend over the same dates.

When an approved draft escapes with a stale claim, broken snippet, or duplicate
topic, preserve it as a regression case. The system improves because failures
become reusable tests—not because the agent receives broader permissions.

See EvalGate's [agent authoring guide](https://www.evalgate.com/docs/guides/write-evals-with-your-agent)
and [runtime controls](https://www.evalgate.com/docs/platform/runtime-controls).