Defining success criteria for agents in Agentic AI - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

What best defines a success criterion for an AI agent?

Imagine you have a robot that cleans your house. What is the best way to describe its success criterion?

A The robot makes a sound when it starts cleaning.

B The robot moves around without bumping into walls.

C The robot cleans every room in the house within 30 minutes.

D The robot has a battery that lasts 2 hours.

💡 Hint

Think about what shows the robot did its main job well.

A success criterion measures if the agent achieved its main goal. Cleaning every room within 30 minutes shows success. Moving without bumping or battery life are important but not the main goal. Making a sound is just a feature.

❓ Metrics

intermediate

Which metric best measures success for a delivery drone agent?

A delivery drone must drop packages at correct locations on time. Which metric best shows if it succeeded?

A Percentage of packages delivered to the correct location within the scheduled time.

B Average battery usage during flights.

C Number of flights completed per day.

D Total distance flown by the drone.

💡 Hint

Focus on the main goal: delivering packages correctly and on time.

The success metric must reflect the agent's main task. Delivering packages correctly and on time is the goal, so percentage of on-time correct deliveries is best. Battery usage, flights per day, and distance flown are secondary factors.

❓ Model Choice

advanced

Choosing a success criterion for a reinforcement learning agent

You train a reinforcement learning agent to play a game. Which success criterion is best to decide if training worked?

A The agent's average score over 100 games exceeds a set threshold.

B The agent's training loss decreases steadily during training.

C The agent's model size is smaller than 10MB.

D The agent's action distribution matches the training data.

💡 Hint

Think about what shows the agent plays the game well, not just training behavior.

Success means the agent performs well in the real task. Average score over many games shows actual performance. Training loss and action distribution relate to learning but don't guarantee good play. Model size is unrelated to success.

🔧 Debug

advanced

Why does this success criterion fail to measure agent success?

An agent's success criterion is: 'Agent completes task if total steps taken is less than 50'. The agent finishes tasks but often takes 60 steps. Why is this criterion problematic?

A It requires the agent to take more than 50 steps to succeed.

B It measures the agent's speed, which is always irrelevant.

C It uses total steps, which is not a measurable quantity.

D It ignores whether the task was actually completed correctly.

💡 Hint

Think about what the criterion checks versus what the agent actually does.

The criterion only checks if the agent is fast (less than 50 steps), but does not check if the task was done correctly. The agent may finish correctly but slowly, so the criterion misses success.

❓ Hyperparameter

expert

How does changing the success threshold affect agent training?

An agent's success criterion is reaching a score above a threshold. What happens if you set this threshold too high during training?

A The agent will learn faster because the goal is more challenging.

B The agent may never reach success, causing training to stall or fail.

C The agent ignores the threshold and trains normally.

D The agent's model size will automatically increase.

💡 Hint

Consider what happens if the goal is impossible or too hard to reach.

If the success threshold is too high, the agent may never achieve success, so training signals become weak or absent. This can cause training to stall or fail. Harder goals don't always speed learning. Model size is unrelated.