Testing Pyramid for AI Agent Systems

The classic testing pyramid doesn't work for AI agents. Here's what does:

**Base: Contract tests (40%)** Verify that tool calls produce valid JSON, API responses match schemas, and LLM outputs parse correctly. These are fast and deterministic.

**Middle: Scenario tests (35%)** End-to-end workflows with mocked LLM responses. "Given this user request, the agent should call tools A then B then respond with X." Use snapshot testing for LLM outputs.

**Top: Evaluation tests (25%)** Run real LLM calls against a test suite. Grade outputs on correctness, helpfulness, and safety. These are slow and non-deterministic — run them nightly, not on every PR.

**Key insight:** Never mock the tool layer. The #1 source of agent bugs is tool call/response mismatches. Always test with real tool schemas.