Babysitter, Auditor, Prayer. Or Tests.

The article argues that reliable LLM agents need code-level tests and assertions, not just prompts or human oversight.

It criticizes prompt-driven control flow and says the better approach is to verify model outputs before any next step or tool action runs.

Practical safeguards include schema checks, predicates, evals, dry-runs, and runtime assertions.

The core idea is to treat LLM agent reliability as a software testing problem, blocking bad outputs before they can cause damage.