Adding Human-in-the-Loop Checkpoints to Your Agent

End-of-lesson check

15 questions · take it digitally for instant feedback at tendril.neural-forge.io/learn/quiz/end-agentic-human-in-loop-r12a1-creators

1. A company is designing an AI agent to help with daily operations. Which type of action should ALWAYS require explicit human approval before the agent executes it?

   1. Generating a summary of yesterday's sales data
   2. Creating a to-do list for the team
   3. Reading internal documentation to find answers
   4. Sending an email to external customers

2. What does the term 'blast-radius' describe in agent safety design?

   1. The memory capacity needed to run the agent
   2. The number of team members who can access the agent
   3. The potential extent of damage if an action is executed incorrectly
   4. The physical size of the computer processing the agent

3. Why might a human-in-the-loop approval system become ineffective if too many checkpoints are added?

   1. Humans will start carefully evaluating each request
   2. The agent will require too much memory to function
   3. Humans will begin rubber-stamping approvals without real scrutiny
   4. The agent will run too slowly and frustrate users

4. A marketing team uses an AI agent to post on social media. The agent wants to publish a promotional tweet. What should happen first?

   1. Show the proposed tweet and require human approval before posting
   2. Post but immediately delete if anyone complains
   3. Post immediately since it's just one tweet
   4. Post only during business hours without approval

5. Which statement best explains why AI cannot fully judge the risk of its own actions?

   1. AI always makes perfect decisions without human help
   2. AI systems are not intelligent enough to understand risk
   3. AI lacks knowledge of the specific business context and priorities
   4. AI should not be used for any business tasks

6. Before requesting human approval for an action, what information should an AI agent provide?

   1. Just the action the agent wants to take
   2. Only technical details about how the action works
   3. Nothing—the human should already know
   4. A clear explanation of what will happen and any relevant details

7. A company configures an AI agent to make small purchases under $10 without human approval. Why might this still be risky?

   1. The agent will buy things the company doesn't need
   2. Small purchases can still violate company policies or laws
   3. The agent will run out of money quickly
   4. There is no risk with small purchases

8. What is the main purpose of requiring human approval BEFORE an agent takes an action, rather than after?

   1. To give the human something to do
   2. To prevent irreversible damage rather than just reporting it after
   3. To satisfy legal requirements only
   4. To make the agent run faster

9. An agent needs to modify pricing for 10,000 products in a database. What is the safest approach?

   1. Show a sample of the changes and require approval before proceeding
   2. Ask the human to manually update each product
   3. Make the changes but keep a backup so it's technically reversible
   4. Make all changes automatically since it's just updating numbers

10. What distinguishes 'confirmation' from 'approval' in human-in-the-loop checkpoint design?

    1. Confirmation is for low-risk actions; approval is for high-risk ones
    2. Confirmation happens after action; approval happens before
    3. There is no meaningful distinction
    4. They mean the same thing in agent design

11. A development team adds a checkpoint requiring human approval every time the agent wants to run a single test. What is likely to happen?

    1. Tests will run faster
    2. The agent will refuse to run any tests
    3. Humans will start rubber-stamping approvals without careful review
    4. Tests will run more accurately

12. Why should an agent pause and request approval before sending external messages?

    1. External messages are technically difficult to send
    2. The recipient might be on vacation
    3. Messages create external-facing impact that cannot be recalled, with potential reputation or legal consequences
    4. External messages cost more money

13. What makes an action 'irreversible' in the context of AI agent design?

    1. The action cannot be easily undone or its effects cannot be recalled
    2. The action requires special computer hardware
    3. The action takes a long time to complete
    4. The action was performed by an AI rather than a human

14. A team configures an agent to spend company money on advertising. What checkpoint approach balances safety with efficiency?

    1. Never let the agent spend any money
    2. Let the agent spend freely up to a small limit with approval required above that threshold
    3. Let the agent spend money only on weekends
    4. Require approval for every purchase regardless of amount

15. What should an AI agent present when asking a human to approve a potentially destructive action?

    1. A clear summary or 'diff' showing exactly what will change
    2. Nothing—the human should ask for details
    3. A list of all possible actions the agent could take
    4. A technical log of all system calls