How do you test AI agent behavior systematically?

[JSONFIRST]

Testing LLM-based agents is hard because outputs aren't deterministic. A few approaches that help:

1. Test on structured intent, not raw text — if your agent acts on JDON instead of raw prompts, you can write deterministic unit tests
2. Confidence thresholds — reject low-confidence inputs before they reach the agent
3. Audit trails — log every JDON object processed

How do you test your agents? Any frameworks or patterns that work well?