Defining success criteria for agents in Agentic AI - Model Metrics & Evaluation

Metrics & Evaluation - Defining success criteria for agents

Which metric matters for this concept and WHY

When we want to know if an agent is successful, we need clear ways to measure it. Success criteria depend on what the agent is supposed to do. For example, if an agent answers questions, accuracy (how many answers are right) matters. If it completes tasks quickly, speed or efficiency matters. Sometimes, we combine several metrics like accuracy, speed, and user satisfaction to get a full picture. Choosing the right metric helps us know if the agent is doing a good job or needs improvement.

Confusion matrix or equivalent visualization (ASCII)

For agents that classify or decide, a confusion matrix helps us see how well they perform. It shows how many times the agent was right or wrong in different ways.

      Confusion Matrix:

          | Predicted Yes | Predicted No
      -----------------------------------
      Actual Yes |     TP       |     FN
      Actual No  |     FP       |     TN

      TP = True Positive (agent correct yes)
      FP = False Positive (agent wrong yes)
      TN = True Negative (agent correct no)
      FN = False Negative (agent wrong no)
    

This helps calculate precision, recall, and accuracy to understand success.

Precision vs Recall tradeoff with concrete examples

Imagine an agent that detects spam emails. If it marks too many good emails as spam (high false positives), users get annoyed. That means precision is low. If it misses many spam emails (high false negatives), spam floods inboxes, so recall is low.

We must balance precision and recall depending on what matters more. For spam, high precision is important to avoid losing good emails. For a medical agent detecting disease, high recall is key to catch all sick patients, even if some healthy ones get flagged.

What "good" vs "bad" metric values look like for this use case

Good success criteria mean the agent meets the goal well. For example:

• Accuracy above 90% for classification tasks.
• Precision and recall both above 85% for balanced detection tasks.
• Low task completion time for efficiency-focused agents.
• User satisfaction scores above 4 out of 5 for interactive agents.

Bad values are low accuracy (below 70%), big gaps between precision and recall, slow responses, or poor user feedback. These show the agent is not successful.

Metrics pitfalls

• Accuracy paradox: High accuracy can be misleading if data is unbalanced. For example, if 95% of emails are not spam, an agent that always says "not spam" has 95% accuracy but is useless.
• Data leakage: When the agent learns from information it should not have, making metrics look better than reality.
• Overfitting indicators: Very high training success but poor real-world results means the agent memorized data instead of learning general rules.
• Ignoring context: Using wrong metrics for the task can hide problems. For example, using accuracy alone for rare event detection.

Self-check question

Your agent has 98% accuracy but only 12% recall on detecting fraud. Is it good for production? Why or why not?

Answer: No, it is not good. The low recall means the agent misses most fraud cases, which is very risky. Even though accuracy is high, it mostly predicts "no fraud" correctly because fraud is rare. For fraud detection, catching as many fraud cases as possible (high recall) is more important.

Key Result

Choosing the right success metric depends on the agent's goal; balancing precision and recall is key for reliable performance.