What problem does it solve?
Once an AI agent is deployed, teams struggle to know which production evaluations and guardrails to set up, and risk creating duplicate, blocking, or misconfigured policies. This Skill grounds recommendations in the agent's actual source code and real ingested traffic, then safely applies only confirmed, non-blocking configurations via the gcx CLI.
Core Features & Use Cases
- Evidence-based recommendations: Reads the agent's system prompt, tools, and entrypoint (cited as file:line) and samples live conversations and generations via
gcx agento11y to decide which eval rules and guards are actually needed.
- Duplicate detection: Inventories existing evaluators, rules, and guards and compares by semantic equivalence (surface, target, intent, action) so nothing is created twice.
- Safe-by-default guardrails: Drafts all new guards in
warn mode (never deny), starts online rules at a low sample_rate, and requires explicit confirmation of the target stack before any write.
- Use Case: Your support agent is live in production and you want to catch quality regressions and redact PII pasted into prompts. The Skill samples a week of traffic, recommends a groundedness rule and a preflight redact guard, drafts reviewable YAML, and applies each only after your explicit yes.
Quick Start
Ask the agent to set up production evaluation and guardrails for your deployed agent, for example: "My agent is in prod — set up online evals and guards for it with gcx."
