The "testing and debugging tools" line in your AI agent platform checklist is the part that hits closest to real production pain. We built pytest-conversational for the same reason: when an AI agent handles 8 turns of multi-step support flow, manual QA can't reliably catch when turn 5 stops following the deterministic part of the workflow because the LLM reasoning step started leaking past the controlled boundary. The approach we landed on: keep the assertions deterministic (rule-based matchers on turn N response shape, role-based permissions, expected tool calls), and let only the agent reasoning be probabilistic. So tests can express things like "if user says X at turn 3, system MUST call tool Y with parameter Z extracted from turn 1, and MUST NOT skip the human handoff trigger at turn 5". Zero LLM in the test side - fully reproducible, runnable in CI without token costs. What surprised us: the more flexible the agent (Hexabot pattern, structured workflow + AI reasoning), the more important determinism on the test side becomes. If both the agent AND the test suite use AI to evaluate output, you have two non-deterministic systems judging each other and the failure modes compound silently. Question: how do you handle workflow versioning for tests? If a tool changes signature between versions of the workflow, do test fixtures auto-detect mismatch or do you rely on integration tests catching the drift after deploy?