Uncle Aaroh by Qambar Raza Book a call
Free scorecard, two minutes

Is your AI agent being trusted on its own word?

12 yes-or-no questions taken from the AI Agent Testing Checklist. Answer for one agent you run or are about to ship. You'll get a score and your biggest gaps straight away.

Your answers stay in your browser. Nothing is saved or sent anywhere.

Boundaries1. Is the agent's goal written down in one sentence, with a definition of done a human can check?
Boundaries2. Are irreversible actions (pay, send, delete) behind an approval step the agent can't skip?
Boundaries3. Does a run fail if it goes over its step, time or token budget, even when the answer is right?
Tool use4. For every action the agent claims ("I searched", "I booked"), is there a tool-call record with arguments and response?
Tool use5. Do you test what happens when a tool returns an error, and check the agent's behaviour changes?
The check step6. Is success decided by a check the agent didn't write and can't edit?
The check step7. Do your tests ignore the agent's "done" message and assert on the actual end state?
The check step8. Would you catch the agent editing or weakening tests to make them pass?
Failure modes9. Have you tested a page or document containing hidden instructions, and confirmed the agent doesn't follow them?
Failure modes10. When the goal is ambiguous or information is missing, does the agent ask rather than guess?
Evaluation11. Do exact checks (state, counts, values) run before any LLM judge scores the output?
Evidence12. Does every run keep its trace, tool-call log, token count and final state long enough to explain it later?
0 of 12 answered
Your result
0

Your biggest gaps