How Agents Go Wrong

End-of-lesson check

15 questions · take it digitally for instant feedback at tendril.neural-forge.io/learn/quiz/end-builders-agentic-agent-failures

1. Which sentence best captures the main idea of 'How Agents Go Wrong'?

   1. Agents should always run without limits or oversight
   2. Agents and chatbots are the same thing in every way
   3. Tools and goals are unnecessary for agent design
   4. Agents fail in funny and scary ways — booking the wrong flight, sending wrong emails, running up bills..

2. Which of the following is part of 'A real failure'?

   1. Always run with no oversight
   2. Disable safety checks for speed
   3. Hide tool calls from the operator
   4. You: 'Email Mom about the party.' Agent emails 'Mom' from your contacts — your roommate's mom, also called Mom in your phone.

3. Which of the following is part of 'Three common failure modes'?

   1. Never log what the agent did
   2. Disable safety checks for speed
   3. Misinterpreting ambiguous instructions
   4. Hide tool calls from the operator

4. Which of the following is part of 'Review date'?

   1. Skip every form of evaluation
   2. Run unbounded retries on any error
   3. Reviewed in 2026. Treat fast-changing product names, prices, availability, and policy details as examples to verify before use.
   4. Avoid taking any actions in the world

5. What is 'ambiguity' in this context?

   1. A trick to bypass approvals
   2. A way to disable the agent's tools
   3. A reason to skip all logging
   4. A core concept covered in How Agents Go Wrong

6. What is 'overconfidence' in this context?

   1. A core concept covered in How Agents Go Wrong
   2. A way to disable the agent's tools
   3. A reason to skip all logging
   4. A trick to bypass approvals

7. What is 'guardrails' in this context?

   1. A trick to bypass approvals
   2. A reason to skip all logging
   3. A way to disable the agent's tools
   4. A core concept covered in How Agents Go Wrong

8. Which is the most informative thing to look at after an agent failure?

   1. Just the user message
   2. The model's release notes
   3. The full trace: prompts, tool calls, returned errors, and decisions
   4. Only the final output

9. Which budget control most directly prevents runaway costs from an agent loop?

   1. A bigger model
   2. A friendly system prompt
   3. A longer context window
   4. A hard cap on steps, tokens, or dollars per task

10. Why does an AI agent need 'tools' such as a browser, calendar, or code runner?

    1. Tools shrink the context window
    2. Tools let the agent take actions in the world instead of only producing text
    3. Tools replace the need for any prompts
    4. Tools make the model speak more naturally

11. What is the best response when an agent suggests an action you do not understand?

    1. Reject everything and stop using the agent
    2. Ask the agent to explain the action and its expected effect before approving
    3. Approve it to keep things moving
    4. Run it twice to be sure

12. Why does a multi-agent system sometimes outperform a single agent on complex jobs?

    1. Multiple agents always cost less
    2. Specialized roles can divide work and check each other
    3. More agents always means more accuracy
    4. Single agents cannot use tools

13. Why is logging every tool call an agent makes a baseline requirement?

    1. Logs are needed to debug, audit, and explain agent behavior to users
    2. Logs replace the need for testing
    3. Logs are only for legal teams
    4. Logs make the model run faster

14. What does an 'eval' for an agent measure?

    1. The exact wording of every prompt
    2. The temperature setting
    3. Whether the agent reliably completes a defined task end to end
    4. How polite the model sounds

15. Why are clear success criteria critical when building an agent?

    1. They make the agent sound smarter
    2. Without them you cannot tell whether the agent worked or guess
    3. They are required by law
    4. They reduce the number of tokens used