These four tests verify that your agent can't be hijacked or manipulated by adversarial input. Run each test multiple times with different phrasings — an agent that fails any of these has a critical vulnerability.

These tests ensure your agent knows the limits of its knowledge and communicates uncertainty appropriately. Hallucination is the most common cause of AI agent production failures — from Air Canada's chatbot inventing refund policies to legal cases citing fake court cases.

LLMs are probabilistic — the same input can produce different outputs. For decision-making agents, this variance is dangerous. Amazon's hiring algorithm was biased. Zillow's algorithm bought houses at the wrong price. Inconsistency is more insidious than a single bad decision — it undermines trust and fairness.

The most dangerous agents are those that exceed their authority without recognizing it. Chevrolet's chatbot agreed to sell cars at losses. Replit's agent deleted production databases. Cigna's system auto-denied medical claims with no escalation threshold. These weren't code bugs — they were behavioral failures where the agent didn't know its own boundaries.

All 15 items must pass. No exceptions. A single failing test means a behavioral vulnerability that users will find. The checklist exists specifically to catch what code reviews miss.

If an item fails, you have two options: (1) Fix the agent behavior, or (2) Reduce the agent's scope so the failure no longer applies. Don't ship with known vulnerabilities hoping they won't be triggered.