Key details for this exam, checked against the published exam outline
Each question shows the correct answer and an explanation of why it is right
A development team is integrating an autonomous coding agent into their SDLC. The agent is asked to refactor a payment processing module. The team wants to ensure the agent's planning phase is clearly separated from its execution phase so that a reviewer can validate the approach before any code changes occur.
Which configuration best enforces this separation?
The correct approach is to have the agent produce a structured plan as a distinct, inspectable artifact, then gate execution behind explicit approval. This enforces the boundary between planning, reasoning, and action, and prevents action until the plan is checked and approved. The other options collapse planning and execution together or remove planning oversight entirely, which defeats the purpose of separating these phases and increases risk of unreviewed changes.
An organization is configuring a coding agent to use a GitHub remote MCP server so it can query issues and pull requests across several repositories. The security team wants to ensure the agent cannot call arbitrary or unvetted tools exposed by the MCP server.
Which action should the team take to restrict the agent's tool usage to only vetted, approved operations?
MCP allow lists let administrators explicitly restrict which tools or operations an agent can invoke through an MCP server, enforcing least-privilege access. Granting broad admin credentials or relying on defaults does not restrict scope, and disabling logging removes observability rather than restricting tool usage, which is counter to security best practices.
A long-running autonomous agent is executing a multi-step migration task across several sessions. Midway through, the agent begins referencing outdated file paths and decisions that were already superseded, causing it to repeat work and produce conflicting outputs.
What is the most likely cause of this behavior, and what should the developer implement to address it?
Repeating superseded work and referencing outdated decisions is a classic symptom of context drift during extended execution. The correct remedy is to detect drift and resynchronize the agent with the current, durable state (e.g., a persisted task log or state artifact) before it resumes. Disabling memory, expanding permissions, or increasing autonomy do not address the root cause of stale or diverging context, and drift does not self-correct without intervention.
After running an evaluation suite against a coding agent's recent pull requests, an engineer reviews the traces and notices the agent successfully understood requirements and produced a correct plan, but repeatedly called the wrong API endpoint when interacting with an internal service, causing failed builds.
How should this failure be classified during root cause analysis?
Because the plan and understanding of the task were correct, but the execution of the tool call (using the wrong API endpoint) was flawed, this is a tool misuse error rather than a reasoning error. Reasoning errors involve flawed understanding or planning, context/environment errors involve stale or missing situational data, and this scenario doesn't involve multiple agents, ruling out an orchestration failure.
In a multi-agent workflow, two agents are each assigned to modify different files within the same feature branch to speed up delivery. During a routine integration check, the orchestration layer detects that both agents have modified an overlapping section of a shared configuration file, producing contradictory changes.
Which action best reflects proper multi-agent conflict detection and resolution practice?
Proper multi-agent orchestration requires detecting overlapping or contradictory changes and routing them for reconciliation, potentially involving human-in-the-loop review, rather than automatically merging, ignoring, or arbitrarily discarding work. This preserves correctness and auditability. Concatenating files or randomly discarding changes risks introducing broken or lost work, and terminating an agent without preserving audit trails violates traceability and accountability requirements.
106 questions covering all exam domains, starting from $20
Exam domains verified against: Official Microsoft GH-600 exam guide, last checked September 2026.
Integrate agents into the software development lifecycle, identifying steps for agents to perform and defining inputs, outputs, and success criteria. Configure planning and reasoning to be distinct from execution, validate plans, and prevent action until an agent has checked and approved its next steps.
Sample question from this domain above: Q1
Select and configure agent tools and Model Context Protocol servers. Integrate agents within development environments by evaluating execution context, scoping agents to specific repositories, and configuring them for CI workflows. Implement error handling, retries, rollbacks, and escalation paths.
Sample question from this domain above: Q2
Implement agent memory strategies by choosing between short-term, long-term, and external memory, then scope memory to task-relevant information. Persist agent state as durable artifacts to allow work resumption without repeating steps. Detect and correct drift during extended execution.
Sample question from this domain above: Q3
Define success criteria and evaluation signals for agent tasks by specifying expected outcomes and operational constraints. Analyze failures by using logs, plans, and traces to identify root causes. Revise instructions, workflows, and constraints based on evaluation results.
Sample question from this domain above: Q4
Apply orchestration patterns to coordinate multiple agents and configure isolation for parallel execution. Detect and resolve conflicts including overlapping code changes and duplicated effort. Implement recovery patterns including rollback and human-in-the-loop when agents fail or show degraded behavior.
Sample question from this domain above: Q5
Classify agent actions by operational, security, and compliance risk to right-size human interventions. Block actions that violate security, compliance, or Responsible AI policies. Scope permissions to enforce least-privilege access and require explicit authorization for irreversible or compliance-sensitive changes.
Common questions about the exam itself
