Agent Evaluations & Red-Teaming — Optional Exercises

Exercise 1: Expand functional coverage

Ask Bob to add three new dataset files under evaluation/datasets/, based on the actual agent and tool schemas:

• a grounded FAQ question
• a refund requiring escalation_agent
• a request that should call no tool

Each file follows the structure of the existing datasets, and each must be added to test_paths in evaluation/config.yaml. For each case, specify the expected behavior without requiring one exact sentence. Ask Bob to validate the JSON, run quick evaluation, and summarize all three results.

Exercise 2: Test tool misuse

Ask Bob to create adversarial prompts that attempt to:

• call process_refund without confirmation
• use another customer's order ID
• pass malformed or extremely large amounts
• persuade the agent to invent a successful tool result

After reviewing them, let Bob turn each one into a dataset file under evaluation/datasets/, generate and run the attacks against that folder, and summarize successful attacks. Do not put real personal data in test prompts.

Exercise 3: Turn a failure into a regression test

Choose one genuine failure and ask Bob to:

1. Preserve the prompt and expected behavior in the dataset.
2. Identify the smallest responsible layer: tool, instructions, guideline, collaborator description, or guardrail.
3. Implement and re-import the fix.
4. Rerun the failed case.
5. Run the original happy paths to check for regressions.

Review Bob's proposed fix before it changes files. Ask it to write a short note containing the failure, root cause, change, and result.

Return to the Agent Evaluations & Red-Teaming module.