Five questions to ask before you trust your agent's output.

Before you ship what an agent wrote, ask:

1. What did it change that I didn’t ask for? Scope creep is silent. 2. What does it assume about the environment? Secrets, network, versions. 3. What happens when the input is hostile? The demo input is never the attacker’s. 4. What’s the blast radius if this is wrong? One record or every record? 5. Can I re-run the check, or am I trusting a screenshot? If you can’t reproduce it, you don’t know it.