AI Zone

Thoughts about AI Development


Project maintained by ITConnector.nl

Guardrails

The guardrail layer

If agents write the code, the useful work is no longer typing. It is deciding what they are allowed to do.

Before you let an intern-with-infinite-energy touch production, you need a layer that does not depend on a human catching every mistake in the moment:

  • architecture rules: where code may live, what may talk to what
  • security defaults: secrets, auth, data handling
  • business invariants: what must never change without a person
  • coding standards that are machine-checkable, not a wiki page
  • no automatic deployment without approvals from end-2-end tests - even if the agent says it is really really good

The point is not to slow the agent down. The point is to make the cheap path the safe path.

A prompt that says “please follow best practices” is not a guardrail. A CI job that rejects the change is.

On the project level you can create a file called AGENT.MD, that you can use to request tracability (e.g. with instructions such as: “When processing a reqest, Log the activity id, category, title, summary, start, end, duration), write the start time as soon as you start with the request in a file called AGENTLOG.csv”

If you have a repo with specific architecture styles are documentation you can use this.

guardrails library