The Eval Set Builder

Help me build an evaluation set for this agent.

Agent purpose: [DESCRIBE]
What good looks like: [DESCRIBE]
Available real data: [DESCRIBE, OR "none yet"]

Produce:
1. Eval categories: what dimensions need measuring
2. For each category, the cases to include:
   - Happy path, representative of typical use
   - Edge cases at boundaries
   - Adversarial cases designed to break it
   - Cases where the correct answer is "I cannot do this"
   - Cases where the correct answer is "I need to ask a human"
3. For each case: input, expected behaviour, how to score it
4. Which cases can be scored automatically vs need human review
5. How many cases are enough to detect a regression
6. What NOT to include — cases that measure the model rather than
   my system

Then address the hard part:
7. How do I score an agent where multiple different action
   sequences are all correct?
8. What do I measure besides final output correctness — tool call
   efficiency, cost, latency, unnecessary actions?
9. What signals a regression versus normal non-determinism?

Be concrete. Give me actual test cases, not categories of test
case.