AI-500 Practice Test

Sample AI-500 Practice Questions

10 real questions from the free AI-500bank. Pick an answer in your head, then open “Show answer” for the explanation.

1. 1. An architect is designing lateral movement prevention for a multiagent solution where a compromised agent must not be able to escalate access to resources outside its own task. Which combination of controls, drawn from Microsoft's agent identity and orchestration security guidance, best achieves this?

   • A. Grant agents in the solution a broadly scoped identity so security reviews happen less often
   • B. Rely exclusively on network-layer firewalls, since identity-layer controls are irrelevant to lateral movement
   • C. Disable RBAC checks between agent-to-agent calls so orchestration isn't slowed down by authorization
   • D. Scope each agent identity's RBAC role assignments narrowly to the resource or resource group it needs, following the principle of least privilege
   Show answer

   Correct answer: D

   Both the agent identity guidance and the multiagent security guidance converge on narrow, resource- or resource-group-scoped RBAC per agent identity, following least privilege, as the control that limits what a compromised agent can reach. A recreates a shared blast radius that enables lateral movement. B ignores that identity-layer least privilege is the guidance's named control, not merely network controls. C removes the exact authorization check that prevents lateral movement.

2. 2. A financial services agent must call a Claude model from Anthropic for a specific reasoning task, while other agents in the same solution use Azure OpenAI models sold directly by Microsoft. Per Foundry's model catalog structure, what is the key operational difference the architect must account for?

   • A. Claude models cannot be deployed through Microsoft Foundry under any circumstances
   • B. Models from partners and community typically cost less than models sold by Azure
   • C. Models sold by Azure are the recommended default; partner models are unsupported within agents
   • D. The Claude model is a "model from partners and community," supported by its provider rather than by Microsoft, unlike models sold by Azure
   Show answer

   Correct answer: D

   Foundry's catalog splits models into "sold by Azure" (Microsoft-supported, deeply integrated) and "from partners and community" such as Anthropic's Claude family (supported by the provider, with Microsoft Support available as help). A is false - the model overview page documents Claude models as available through Foundry. B is an unfounded pricing generalization not stated in the documentation. C is false - nothing in the catalog restricts partner models from being used within agents.

3. 3. An architect reviews Application Insights performance data for a hosted agent and finds sustained CPU peaks around 90% of the allocated 1 vCPU / 2 GiB sandbox tier during representative load testing, while memory usage stays well below its allocation. Per Foundry's right-sizing guidance, what should the architect do for the next agent version?

   • A. Raise the CPU allocation to the next tier, since sustained peaks exceeding roughly 70% of allocation indicate the current tier is undersized
   • B. Lower both CPU and memory allocation, since sustained high CPU usage indicates the sandbox is oversized
   • C. Leave the allocation unchanged, since Foundry automatically resizes sandboxes at runtime without creating a new version
   • D. Switch to a shared replica pool instead of per-session sandboxes to reduce CPU pressure
   Show answer

   Correct answer: A

   Foundry's right-sizing guidance states that if sustained peaks exceed roughly 70% of allocation, the next agent version's allocation should be raised; 90% sustained CPU well exceeds that threshold. B contradicts the guidance's meaning of high sustained usage. C is false - sandbox sizing is fixed per immutable agent version and requires deploying a new version to change. D is false - hosted agents scale per session, not via a shared replica pool; there is no replica count to configure.

4. 4. An architect is decomposing a claims-processing goal into agents. The workload has three distinct security boundaries (intake, adjudication, payout), each stage benefits from parallel specialization, and cross-functional problems require agents with different tool access. A single agent with tools is unable to enforce the required security separation. Which factor from Microsoft's multiagent guidance most directly justifies moving to a multiagent decomposition here?

   • A. The desire to reduce the total number of model calls made across the workflow
   • B. The preference for a single accumulated context window shared across the workflow's stages
   • C. The requirement to minimize latency by avoiding any coordination overhead
   • D. The presence of distinct security boundaries per stage, which a single agent cannot enforce
   Show answer

   Correct answer: D

   Microsoft's guidance names distinct security boundaries per stage as a scenario where a single-agent approach becomes infeasible even if it could otherwise handle the task. A is false - multiagent orchestration increases model calls, not reduces them. B is the opposite of what multiagent decomposition provides (each agent gets scoped context). C is false - multiagent patterns add coordination overhead by design, they don't eliminate it.

5. 5. A site-reliability team is designing a remediation workflow for low-risk production incidents. The exact remediation steps aren't known upfront: a manager agent must consult diagnostics, infrastructure, and rollback agents, revise its plan as new information arrives, and maintain a full audit trail for postincident review. Which orchestration pattern should the architect specify?

   • A. Sequential orchestration with a fixed four-stage pipeline
   • B. Magentic orchestration with a manager agent that builds and refines a task ledger
   • C. Concurrent orchestration where specialist agents run in parallel on the same incident
   • D. Group chat orchestration limited to two agents in a maker-checker loop
   Show answer

   Correct answer: B

   Magentic orchestration is designed for open-ended problems with no predetermined solution path, where a manager agent iterates, backtracks, and maintains a task ledger with a full audit trail - exactly this SRE scenario. A requires a known, fixed step order, which this scenario lacks. C doesn't support the iterative plan-building and ledger tracking required. D is designed for validation/consensus, not dynamic incident remediation planning.