Agents Demystified: What They Are and Are Not

End-of-lesson check

15 questions · take it digitally for instant feedback at tendril.neural-forge.io/learn/quiz/end-ai-foundations-agents-final1-creators

1. Which statement best describes the fundamental architecture of an AI agent?

   1. A pre-programmed script that executes the same steps regardless of context
   2. A standalone language model that generates text without any external interaction
   3. A neural network that directly controls robotic hardware without any intermediate processing
   4. A model embedded in a continuous loop that plans, acts, observes results, and adjusts its approach

2. What is the primary mechanism by which AI agents can recover from small mistakes during task execution?

   1. Human operators manually correct each error in real-time before the agent continues
   2. The model automatically retries every action three times before proceeding
   3. The agent has a built-in encyclopedia that prevents it from making factual errors
   4. The loop structure allows the agent to observe the results of each action and adjust its next steps

3. Why do AI agents that work perfectly in demos often fail when deployed to real production environments?

   1. Demos use narrow, curated tools and short tasks, while production requires guardrails, timeouts, error handling, and human interruption capabilities
   2. The AI model degrades after running for more than a few minutes
   3. Production environments lack the internet connection that demo agents rely on
   4. Demo agents use more advanced AI models than production systems

4. In the ReAct loop architecture, what does the 'Act' component represent?

   1. The agent generating a final answer to present to the user
   2. The process of loading the model into memory
   3. The agent's initial prompt setup phase
   4. The execution of a tool based on the model's reasoning

5. What specifically causes AI agents to 'derail' when executing long sequences of steps?

   1. The agent runs out of memory after about 20 steps
   2. Compounding errors — small mistakes early in the task propagate and corrupt later reasoning
   3. The model spontaneously begins ignoring its instructions after too many iterations
   4. External hackers inject errors into the agent's execution pipeline

6. Why is most of the complexity in an AI agent system found in prompts and tool design rather than the loop code itself?

   1. Prompt complexity is an illusion — the actual intelligence resides in the loop logic
   2. The loop code is generated automatically and requires no human design effort
   3. Tool design is irrelevant; agents work the same way regardless of available tools
   4. The loop is conceptually simple (essentially a while-loop), but the quality of reasoning and tool selection depends on how prompts and tools are designed

7. What would happen if an AI agent were deployed without any step budget or time limit?

   1. The agent would automatically improve its accuracy the longer it runs
   2. The agent would eventually achieve perfect task completion given infinite time
   3. The agent might continue executing indefinitely, potentially running up costs or getting stuck in unproductive loops
   4. The agent would shut down gracefully after completing its primary task

8. Which capability is explicitly listed as something AI agents CAN do well in the lesson?

   1. Operate asynchronously over long time horizons without constant human supervision
   2. Avoid all errors by checking every action against a knowledge base
   3. Replace human decision-making in medical and legal domains
   4. Generate perfectly accurate code on the first try every time

9. What is 'observability' in the context of deployed AI agents and why does it matter?

   1. A feature that allows the agent to watch other agents and learn from them
   2. The agent's ability to see and interpret visual inputs like cameras
   3. The agent's capacity to observe its own internal reasoning process
   4. The ability to monitor, log, and understand what the agent is doing at each step to debug failures

10. Why can't AI agents currently replace human oversight in high-stakes domains like healthcare or finance?

    1. The agents' vulnerability to compounding errors means mistakes can cascade in ways that are dangerous in high-stakes contexts
    2. AI models are too slow to make real-time decisions in these fields
    3. These domains have banned AI agents from being used
    4. AI agents cannot process the specialized data formats used in healthcare and finance

11. What is the relationship between an agent's 'autonomy' and its potential for failure?

    1. Agents with high autonomy never fail because they learn from their mistakes
    2. More autonomy means the agent can make more independent decisions, which increases the potential for unchecked errors
    3. Higher autonomy always leads to better outcomes because the agent makes fewer mistakes
    4. Autonomy has no relationship to failure rates in agent systems

12. What distinguishes a simple chain of prompts from an AI agent system?

    1. Agents use a feedback loop that allows them to adjust based on previous results, while chains execute sequentially without observing outcomes
    2. Simple chains can only execute one tool before ending
    3. Simple chains are more advanced than agents because they don't require loops
    4. There is no functional difference; agents are just a marketing term for chains

13. If you were building a minimal AI agent, what would the core code structure look like?

    1. A complex neural network with multiple hidden layers
    2. A simple while-loop that calls the model, parses its response, executes one tool, and feeds the result back until the task is done
    3. A database query system that retrieves pre-written answers
    4. A rigid if-else decision tree with pre-defined responses

14. Why might adding more tools to an AI agent make it MORE likely to fail in unexpected ways?

    1. Tools interfere with each other and cause hardware failures
    2. The agent will automatically select the best tool every time
    3. Additional tools slow down the agent so much it times out
    4. With more tools, the agent has more opportunities to choose the wrong tool or make poor combinations of tool calls, increasing error surface area

15. What feature allows a human to take control back from an AI agent that is behaving unexpectedly?

    1. A human interrupt capability that allows operators to pause or stop the agent mid-execution
    2. A shutdown button built into the model weights
    3. The agent automatically stops when it detects it's making mistakes
    4. A voice command system that the agent always obeys