support@Free-Braindumps.com
Register Login
Cisco
CompTIA
ISC
EC-Council
EXIN
Fortinet
ITIL
Juniper
Microsoft
VMware
Avaya
SAP
Google
NVIDIA
Salesforce
Amazon
Palo Alto Networks
ISC2
CrowdStrike
Anthropic
All Vendors
  1. Home
  2. Exams
  3. Microsoft
  4. AI-500

Free AI-500 Exam Braindumps
Designing and Implementing Multi-Agent AI Solutions Questions 46–50 (page: 11)

Updated On: 2-Oct-2026
Page 11 of 23
PDF with 107 Questions
View Related Case Study

You have a Microsoft Foundry project for a multi-agent coding solution. The project includes agents that call
GitHub tools. The tools return issue numbers, pull request numbers, branch names, commit SHAs, and workflow run links. Handoff payloads store copied acceptance criteria, check results, and repository names.
After long sessions, the review agent comments about outdated pull requests or attributes that failed workflow runs to the wrong branch. Trace logs show truncated tool output and stale check results in memory.
You need to improve entity continuity across agent transitions and reduce context window pressure.
What should you do?

  1. Persist GitHub artifact identifiers in memory and handoffs. Rehydrate state from GitHub on each transition.
  2. Persist the complete tool output in memory and handoffs. Inject the output before each invocation.
  3. Persist a conversation-scoped thread across all the agents. Transfer the full transcript during each handoff.
  4. Persist separate copied summaries for each agent. Reconcile the summaries during the review transition.

Answer(s): A

Explanation:

Persisting GitHub artifact identifiers and rehydrating state from GitHub on each agent transition is the correct and most effective action to resolve this problem.
In long-running LLM multi-agent sessions, passing full tool outputs (like massive JSON payloads of PR details,
check logs, or commit histories) through agent handoffs inevitably leads to context window saturation and truncation. Furthermore, because GitHub repositories change dynamically during a coding session (e.g., a workflow finishes, a new commit is pushed), storing static check results in agent memory guarantees stale state.
Eliminates Truncation: By passing only lightweight entity identifiers (e.g., repository_name, pr_number,
commit_sha), you dramatically reduce the payload size transferred between agents, keeping the context window clean and saving token costs.
Guarantees State Freshness: Instead of relying on a snapshot of what a workflow run was doing 20 minutes ago, the receiving agent queries the live GitHub API to fetch the absolute latest status, preventing the review agent from acting on stale data.
Maintains Entity Continuity: Standardizing an explicit "State Schema" for handoffs ensures every agent in the multi-agent pipeline is aligned on exactly which branch, PR, or commit is currently being targeted.


Reference:

https://techcommunity.microsoft.com/blog/azuredevcommunityblog/building-and-operating-a-microsoft-foundry
-hosted-agent-with-gitops-and-github-t/4523393



View Related Case Study

You have a Microsoft Foundry production multi-agent solution.
After a model and tool update, the 95th percentile (p95) latency service-level indicator (SLI) exceeds four seconds and the tool-call failure SLI exceeds two percent for multiple five-minute windows. Production changes require operator approval.
You need to implement a remediation strategy that creates incidents after only five minutes of SLI violation.
The strategy must meet the following requirements:
● Perform telemetry-and-change triage before mitigation.
● Require human approval before production changes.
● Create a post-incident artifact after resolution.
What should you do?

  1. For Azure Container Apps health and resource limits, create Azure Monitor alerts that have a five-minute window, and then trigger a serial Foundry workflow for triage, approval-gated remediation, and post-incident artifact creation.
  2. For the latency and tool-call failure SLIs, create Azure Monitor alerts that have a 15-minute window, and then trigger a serial Foundry workflow for triage, approval-gated remediation, and post-incident artifact creation.
  3. For the latency and tool-call failure SLIs, create Azure Monitor alerts that have a five-minute window, and then trigger a serial Foundry workflow for triage, approval-gated remediation, and post-incident artifact creation.
  4. For the latency and tool-call failure SLIs, create Azure Monitor alerts that have a five-minute window, and then trigger a concurrent Foundry workflow for triage, approval-gated remediation, and post-incident artifact creation.

Answer(s): C

Explanation:

To resolve this issue, you should create separate Azure Monitor alerts for the latency and tool-call failure SLIs with a five-minute evaluation window, and then trigger a serial Azure DevOps / Foundry pipeline workflow to handle the triage, approval-gated remediation, and post-incident artifact creation.
A serial workflow is required because your three operational requirements are strictly sequential. You cannot safely approve a production change until triage is complete, and you cannot generate an accurate post-incident artifact until the remediation has been applied and verified.


Reference:

https://learn.microsoft.com/en-us/azure/foundry/agents/concepts/tool-best-practice



View Related Case Study

HOTSPOT (Drag and Drop is not supported)
You have a Microsoft Foundry multi-agent refund solution that includes the following agents:
● A policy agent that retrieves refund rules from Microsoft SharePoint Online
● An action agent that issues billing credits through Azure Logic Apps
Users report that some refund responses cite the incorrect policy, and one run attempted to issue a credit after a prompt injection attack.
You need to implement a human review process that meets the following requirements:
● Reviews only higher-risk runs
● Prevents unapproved billing changes
● Gives reviewers enough existing runtime evidence to assess policy grounding, task completion, and tool use
What should you configure? To answer, select the appropriate options in the answer area.
Note: Each correct selection is worth one point.
Hot Area:

  1. See Explanation section for answer.

Answer(s): A

Explanation:



Box 1: Thread logs, run steps, and evaluator results
To give reviewers full visibility into policy grounding, task completion, and tool tracking, ingest runtime metadata from Foundry Agent Service and package them inside the reviewer panel:
Thread Logs (Policy Grounding & Injections): Expose the complete multi-turn conversations session history.
This allows reviewers to trace the exact user inputs—flagging malicious injection payloads—and verify whether the Policy Agent accurately extracted rules from Microsoft SharePoint Online.
Run Steps (Task Completion & Tool Use): Render the sequential agentic execution steps. Reviewers will see precisely when the Policy Agent invoked its tools, how it formulated the internal reasoning loop, and when the
Action Agent was queued to call the Logic App.
Evaluator Results (Risk Assessment): Surface the Microsoft Foundry Risk and Safety Evaluations metrics directly to the UI. Display concrete scores for Indirect Jailbreak / Prompt Injection Vulnerability and semantic text relevance (grounding) to explicitly show why the system flagged the run.
Box 2: Pause credit actions until supervisor approval
To implement a secure human review process in Azure Logic Apps for your Microsoft Foundry multi-agent refund project, you should use the Create Autonomous AI Agentic Workflows feature to add a Wait for approval checkpoint before executing the billing credit action.
Steps
Pause Credit Actions: Configure your Logic Apps workflow to pause immediately after the policy and action agents process a request, using a built-in approval connector that halts execution before the Azure Logic Apps billing API is called.
Filter for Higher-Risk Runs: Add a conditional check in the workflow (such as evaluating risk scores, high refund amounts, or anomaly flags triggered by content safety guards) so that low-risk transactions auto-approve while higher-risk runs route to a supervisor.
Prevent Unapproved Changes: Ensure the billing action branch remains locked and inactive until an authorized reviewer explicitly clicks "Approve" in the notification payload.
Provide Runtime Evidence: Include conversation history, retrieved SharePoint policy snippets, and tool-call audit logs directly inside the approval notification email or task card so reviewers can verify grounding and detect prompt injection attempts.


Reference:

https://learn.microsoft.com/en-us/azure/foundry/agents/concepts/runtime-components



View Related Case Study

HOTSPOT (Drag and Drop is not supported)
You have a Microsoft Foundry multi-agent solution. The solution includes an orchestrator agent that routes requests to a specialist agent.
Occasionally, the specialist agent has the following issues:
● Ignores its assigned scope
● Answers questions it should decline
● Forgets to include the required compliance disclaimer
You need to use the agent optimizer in Foundry Agent Service to improve the specialist agent's prompt behavior automatically. The solution must enable prompt rewriting for the baseline.
What should you configure? To answer, select the appropriate options in the answer area.
Note: Each correct selection is worth one point.
Hot Area:

  1. See Explanation section for answer.

Answer(s): A

Explanation:



Box 1: Instruction tuning
To solve these behavioral issues automatically using the Agent Optimizer in the Foundry Agent Service, you should set up the primary optimization target to instruction.
Purpose: The Instruction target is the default optimization mode in Microsoft Foundry specifically built to analyze where an agent's responses fall short and rewrite the underlying system prompt.
The optimizer runs a closed-loop evaluation.
When the specialist agent fails your defined rubrics (e.g.,
exceeding scope, answering unassigned topics, or omitting compliance texts), the optimizer detects these failures and dynamically infuses alternative system instructions, boundaries, and required formatting rules to overcome them.
Box 2: instructions.md
To enable prompt rewriting (instruction tuning) in the Foundry Agent Service Agent Optimizer, you must include an instructions.md file as your baseline artifact.
The Agent Optimizer looks for specific configurations inside a .agent_configs/baseline/ directory within your project structure. To configure it to automatically address your specialist agent's behavior (ignoring scope,
answering off-topic questions, and missing disclaimers), implement the following setup:


Reference:

https://devblogs.microsoft.com/foundry/agent-optimizer-build2026/



View Related Case Study

DRAG DROP (Drag and Drop is not supported)
You have a Microsoft Foundry multi-agent customer support solution. The solution includes an orchestrator that invokes multiple agents within a single conversation. The agents call Azure OpenAI models and external Model
Context Protocol (MCP) servers.
You need to recommend a monitoring strategy that meets the following requirements:
● Detects unavailable agents on a defined schedule
● Correlates telemetry across the orchestrator, agents, model calls, and external calls for each customer interaction
What should you include in the recommendation? To answer, drag the appropriate strategies to the correct requirements. Each strategy may be used once, more than once, or not at all. You may need to drag the split bar between panes or scroll to view content.
Note: Each correct selection is worth one point.
Select and Place:

  1. See Explanation section for answer.

Answer(s): A

Explanation:





Box 1: Endpoint monitoring
The best strategy to use for this requirement is endpoint monitoring.
Meets the Defined Schedule Requirement: Endpoint monitoring (often referred to as synthetic monitoring or uptime checking) is explicitly designed to ping endpoints, run synthetic transactions, or invoke health-check endpoints on a pre-defined schedule (e.g., every 1 or 5 minutes).
Detects Agent Availability: In a Microsoft Foundry multi-agent system, agents are deployed with a managed endpoint or exposed via agent-to-agent (A2A) endpoints. By scheduling endpoint checks directly against the orchestrator or individual agent availability routes, you can immediately detect if a component is unresponsive or failing before real users encounter the error.
Box 2: Distributed traces with context
The best strategy to use for customer integration diagnostics is distributed traces with context.
End-to-End Correlation: Distributed tracing passes a unique trace identifier (context) across process and network boundaries. This links the orchestrator, individual agents, Azure OpenAI model calls, and external
Model Context Protocol (MCP) servers into a single, cohesive timeline for each specific customer interaction.
Deep Diagnostic Capabilities: When a customer interaction experiences high latency or an error, distributed traces allow you to pinpoint exactly which component caused the breakdown—whether it was a slow response from an external MCP server or an issue within the orchestrator itself.
Asynchronous & Multi-Agent Mapping: Multi-agent architectures involve non-linear, delegated workflows. A
context-aware distributed trace maps out these intricate execution paths clearly.


Reference:

https://learn.microsoft.com/en-us/azure/foundry-classic/agents/how-to/connected-agents



  PDF
Viewing page 11 of 23
Viewing questions 51 - 55 out of 107 questions


Post your Comments and Discuss Microsoft AI-500 exam prep with other Community members:

AI Tutor AI Tutor 👋 I’m here to help!