AI Research Agents: Cool Power, Real Risks

End-of-lesson check

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

1. Which sentence best captures the main idea of 'AI Research Agents: Cool Power, Real Risks'?

   1. Tools and goals are unnecessary for agent design
   2. Agents should always run without limits or oversight
   3. Agents and chatbots are the same thing in every way
   4. Some AI tools (Deep Research, Perplexity Pro) do hours of web research for you in minutes. Powerful — but verify what they bring back.

2. Which of the following is part of 'Some examples'?

   1. Hide tool calls from the operator
   2. Use for: getting a quick overview of a topic you do not know.
   3. Never log what the agent did
   4. Disable safety checks for speed

3. Which of the following is part of 'The rule'?

   1. Skip every form of evaluation
   2. Avoid taking any actions in the world
   3. Approve all actions automatically
   4. Research agents save time. They do NOT replace you reading the actual sources for important stuff. For homework that matters, verify.

4. Which of the following is part of 'You did it!'?

   1. Avoid taking any actions in the world
   2. Great work — you just leveled up.
   3. Skip every form of evaluation
   4. Never log what the agent did

5. What is 'research agents' in this context?

   1. A reason to skip all logging
   2. A trick to bypass approvals
   3. A core concept covered in AI Research Agents: Cool Power, Real Risks
   4. A way to disable the agent's tools

6. What is 'automation' in this context?

   1. A core concept covered in AI Research Agents: Cool Power, Real Risks
   2. A trick to bypass approvals
   3. A way to disable the agent's tools
   4. A reason to skip all logging

7. What is 'verification' in this context?

   1. A core concept covered in AI Research Agents: Cool Power, Real Risks
   2. A reason to skip all logging
   3. A trick to bypass approvals
   4. A way to disable the agent's tools

8. What is the biggest risk of trusting a research agent's report without checking it?

   1. It will always refuse to answer
   2. It can confidently cite sources that do not exist or misread real ones
   3. It might be too short
   4. It can only use one source at a time

9. What is the most reliable way to keep an autonomous agent from going off the rails on a long task?

   1. Trust the model to know when to stop
   2. Run it for as many steps as possible without checking in
   3. Set a clear goal, a step budget, and review checkpoints
   4. Disable its tools so it can only think

10. Which of these is the strongest indicator that an agent workflow is ready to scale?

    1. It passes a repeatable eval, has cost in budget, and a rollback plan
    2. It runs without any logging
    3. It worked once for one user
    4. It used the latest model

11. What is the safest first place to deploy a brand new agent?

    1. Production, against real customers
    2. Inside a critical billing system
    3. On a public server with no auth
    4. A sandbox or low-stakes task with reversible actions

12. Why is it dangerous to give an agent access to your email and calendar without scoped permissions?

    1. Broad access means a single misstep can send the wrong message or wipe events
    2. Scopes only matter for paid accounts
    3. It will refuse to work
    4. Scoped permissions slow the model down

13. Which is the best way to think about an agent's 'autonomy level'?

    1. Autonomy does not affect risk
    2. It is binary — fully autonomous or not
    3. It is a slider from suggestion only, to ask-then-act, to act-then-report, to fully autonomous
    4. It depends only on the model size

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

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

15. Before letting an agent take a destructive action, what is the safest default?

    1. Skip approvals if the user trusts the agent
    2. Approve once and let the agent repeat forever
    3. Require explicit human approval for the specific action
    4. Hide the action from any log