Skip to main content
AI6 min read

AI agent guardrails checklist

A pre-production checklist for agentic features — escalation, evals, logging, and cost controls.

Before you ship

Define allowed tools and data sources; block unscoped file or network access.

Write eval cases for happy path, refusal, and escalation scenarios.

Set token and spend alerts; log prompts and outputs for support review where policy allows.

Config-driven knowledge article — edit in src/config/knowledge.ts until CMS editors are available.