Microsoft · AI-500
Guardrails for multi-agent solutions
AI-500 · Lesson 15 of 16 · 2:07:35
YouTube loads only after you press play. This chapter opens at its timestamp in the full course. Watch on YouTube ↗
In this lesson
Prepare for the AI-500 exam (Microsoft Certified: Multi-Agent AI Solutions Expert). In this lesson: a multi-intervention guardrail strategy across user input, tool calls, tool responses, and outputs, Foundry guardrails (controls, intervention points, actions, agent guardrails in preview, latency), custom guardrails for domain rules with blocklists, classifiers, LLM judges, and retrieval checks, and guardrail testing with a coverage matrix, synthetic data, and false positive and false negative rates. Includes three original practice questions.
Preview the topics
- A partner’s hotel API returns descriptions that sometimes contain hidden instructions aimed at Contoso’s agent. Where should a guardrail scan for this?
- Contoso’s legal team adds a rule: the assistant must never present a visa decision as final, and the wording varies a lot from one answer to the next. Which detection method fits?
- Contoso’s new guardrail blocks nearly every injection attempt in testing, but in production it also blocks one traveler request in five. Which measure did the test suite miss?