The real GH-600 exam is scenario-heavy: it describes a situation involving GitHub Copilot coding agent, MCP servers, or multi-agent workflows, then asks what you'd configure, classify, or approve. These 12 questions are written the same way, and they stay inside one continuous story so the domains build on each other the way a real rollout would — architecture decisions in the first two questions shape the tool, memory, and evaluation problems that show up later.
Want the domain breakdown and weighting first? See the GH-600 study guide.
Missing more than 2 or 3 questions in one domain below is a signal to spend more time there before booking — see the full GH-600 domain list and sample question for the official weighting.
1. Prepare agent architecture and SDLC processes
Questions 1–2 · Northwind Ledger's rollout continues
1Prepare agent architecture and SDLC processesNorthwind Ledger, a fintech engineering org, is piloting GitHub Copilot coding agent on its billing-service repository. The platform lead wants the agent to produce a written plan of the files it intends to touch and the changes it intends to make, and wants that plan reviewed before any code is written.Which configuration goal does this represent?Tap to reveal the four choices and the correct answer →
This is a direct instance of the architecture domain's core skill: separating planning from execution and blocking action until a human (or an automated check) has validated the plan. It reduces the blast radius of a wrong assumption before any file is touched.
A: This is the opposite of what the platform lead is asking for — skipping planning removes the review step entirely.
C: Codespaces is a development environment choice, unrelated to whether planning and execution are separated.
D: Branch protection controls who can merge code — it does not create or enforce a plan-then-approve workflow before an agent acts.
2Prepare agent architecture and SDLC processesTwo weeks into the pilot, a Northwind Ledger engineer notices the coding agent tends to rewrite unrelated formatting across a file whenever asked to fix one function — a pattern the team has seen cause noisy, hard-to-review diffs on three separate tasks now.What should the platform lead do first, per the architecture domain's guidance?Tap to reveal the four choices and the correct answer →
The architecture domain explicitly includes identifying and mitigating common anti-patterns in agents as part of integrating them into the SDLC — a recurring, identifiable bad habit like scope creep in diffs is exactly what this sub-skill targets, typically by constraining the agent's instructions or output scope.
B: A full shutdown is disproportionate to a known, mitigable pattern and throws away the pilot's progress.
C: Noisy diffs slow down review and hide the actual change — treating them as harmless ignores real cost.
D: This is a planning/scope issue, not a memory or state issue — waiting for the wrong domain's review delays a fix that belongs here.
2. Implement tool use and environment interaction
Questions 3–4 · Northwind Ledger's rollout continues
3Implement tool use and environment interactionNorthwind Ledger wants its coding agent to query a read-only reporting database for account balance formats while it works. A developer sets this up by launching a local process with a command and arguments that starts an MCP server pointed at the database connection string.What kind of MCP server has the developer configured?Tap to reveal the four choices and the correct answer →
Launching an MCP server by specifying a command (like npx) and its arguments, rather than pointing to a hosted HTTPS endpoint, is the local MCP server pattern — the server process runs on the machine executing the agent, not on a remote host.
A: OAuth authentication is specific to GitHub's remote, HTTP-based MCP server — this setup runs a local process instead.
C: api.githubcopilot.com/mcp/ is GitHub's own hosted remote MCP endpoint — this scenario describes a self-launched local process against a different database.
D: A registry is a list of approved MCP servers for allowlist enforcement, not the server configuration itself.
4Implement tool use and environment interactionThe reporting database MCP server needs a connection string that includes a password. Northwind Ledger's security team wants this value available to the agent's MCP configuration without ever appearing in the workflow YAML or being readable by unrelated workflows.How should this value be stored so the MCP configuration can use it?Tap to reveal the four choices and the correct answer →
GitHub's own documentation states that only variables and secrets with names prefixed COPILOT_MCP_ are made available to the MCP configuration for Copilot's coding agent — this is the specific, documented mechanism for this exact need.
A: A plain, unprefixed repository variable is not automatically exposed to the MCP configuration and offers no secret masking.
B: Hardcoding a password directly into YAML exposes it to anyone who can read the repository.
D: A committed .env file exposes the secret in git history even if later removed, and is not the mechanism GitHub documents for this.
3. Manage memory, state, and execution
Questions 5–6 · Northwind Ledger's rollout continues
5Manage memory, state, and executionA Northwind Ledger agent task spans several hours: refactor a module, then update its tests, then update its documentation, resuming after a lunch break. Midway through, the engineer notices the agent has started re-reading files it already processed and repeating earlier reasoning steps instead of building on them.What is this symptom an example of, and what does the memory/state domain call for?Tap to reveal the four choices and the correct answer →
Repeating steps or diverging from prior decisions over a long-running task is the textbook description of context drift. The memory/state domain specifically calls for capturing task progress as durable artifacts so an agent can resume without repeating steps, and for detecting and correcting drift as it happens.
A: Tool misuse is a category from the evaluation domain describing incorrect tool usage, not memory continuity loss over time.
C: This symptom is about memory continuity, not a policy violation requiring an approval gate.
D: Repeated steps waste time and can produce inconsistent results — this is precisely the failure mode the domain asks you to detect and fix.
6Manage memory, state, and executionA Northwind Ledger developer asks whether a fact the agent learned about their personal code-review style preferences on one repository should be visible when the same developer works with the agent on a completely different, unrelated repository.Which memory-scoping principle governs whether that preference should carry over?Tap to reveal the four choices and the correct answer →
This mirrors how GitHub's own Copilot memory feature actually separates the two: repository facts stay confined to their origin repository for privacy, while user-level preferences are tied to the individual and can travel with them across repositories where they use Copilot.
B: This describes repository-level facts only — it ignores that user preferences are explicitly scoped to the person, not the repo.
C: Treating personal preferences as global and shared organization-wide would leak individual working style to every other user, which is not how scoping works.
D: Memory scoping follows defined rules (user-level vs. repository-level), not randomness.
4. Perform evaluation, error analysis, and tuning
Questions 7–8 · Northwind Ledger's rollout continues
7Perform evaluation, error analysis, and tuningA Northwind Ledger agent task to 'add input validation to the payments endpoint' comes back with validation logic that technically compiles but checks the wrong field entirely — the agent correctly understood the tools available but misjudged what the code actually needed.Which root-cause category does this failure fall into?Tap to reveal the four choices and the correct answer →
The agent used its tools correctly and had no missing context — it drew the wrong conclusion about what the task required. That is a reasoning error, one of the three root-cause categories the evaluation domain names explicitly (alongside tool misuse and context/environment issues).
A: Tool misuse would mean the agent used a tool incorrectly (wrong parameters, wrong tool for the job) — here the tools worked fine.
B: A context/environment issue would mean the agent lacked information it needed or hit an environment constraint — this was a judgment failure with full context available.
D: Nothing here involves a policy, permission, or compliance boundary — it is a quality-of-reasoning problem.
8Perform evaluation, error analysis, and tuningAfter classifying several agent failures as reasoning errors on ambiguous validation tasks, the Northwind Ledger platform lead wants to reduce the recurrence rate without touching what tools the agent can access.Which tuning action fits a reasoning-error root cause, per the evaluation domain?Tap to reveal the four choices and the correct answer →
The evaluation domain ties each root-cause category to a matching tuning lever: reasoning errors are addressed by revising instructions, workflows, or constraints — giving the agent clearer, less ambiguous direction — rather than by touching tools or memory, which address different root causes.
B: Refining tool access addresses tool-misuse failures, not a reasoning error where the tools were used correctly.
C: Refining memory usage addresses continuity/drift issues, not a one-off misjudgment about task intent.
D: Adding a redundant tool does nothing to correct ambiguous task framing, and increases the tool-use surface the agent must manage.
5. Orchestrate multi-agent coordination
Questions 9–10 · Northwind Ledger's rollout continues
9Orchestrate multi-agent coordinationNorthwind Ledger runs two coding agents in parallel on the same feature branch: one updating the payments module and one updating its test suite. Both agents open commits that modify the same helper file within minutes of each other, and the resulting pull requests contain contradictory changes to that file.What is happening, and what does the multi-agent domain call for?Tap to reveal the four choices and the correct answer →
Overlapping code changes between agents working in parallel is named explicitly in the multi-agent domain as a type of conflict to detect and resolve — the fix is isolating agents appropriately or coordinating their scope, not a memory or guardrails fix.
A: Memory expiration governs how long facts persist, not whether two independent agents step on the same file.
C: Merging two agents into one is not a documented guardrails response, and guardrails governs risk and human oversight, not file-level conflicts.
D: Contradictory changes to the same file will break the build or silently overwrite work — this needs an orchestration response, not indifference.
10Orchestrate multi-agent coordinationThree months into the rollout, Northwind Ledger wants to retire one of its older coding agents (originally configured for a since-decommissioned service) without breaking the audit trail auditors rely on for the workflows it participated in.Which lifecycle practice does the multi-agent domain call for here?Tap to reveal the four choices and the correct answer →
Managing the lifecycle of agents within multi-agent workflows explicitly includes retiring agents while preserving auditability and workflow continuity — the audit trail for past decisions needs to survive the agent that made them.
A: Deleting records defeats the auditability the compliance team depends on for past decisions.
C: Reusing an identity for an unrelated agent would corrupt the historical audit trail by attributing new actions to an old identity.
D: Leaving a decommissioned-service agent active indefinitely is itself a governance risk the lifecycle skill is meant to prevent.
6. Implement guardrails and accountability
Questions 11–12 · Northwind Ledger's rollout continues
11Implement guardrails and accountabilityA Northwind Ledger coding agent opens a pull request that, if merged, would push directly to the default branch and skip the team's usual review. The workflow that would run as a result of this PR does not start automatically.What default guardrail behavior does this describe?Tap to reveal the four choices and the correct answer →
This matches GitHub's own documented default behavior for coding agent guardrails: agents can't push to the default branch or merge PRs by default, and workflows triggered by an agent's PR require write-access approval before running — a repository admin can disable this, but it is on by default.
B: This is the opposite of the documented default — agents start restricted, not with full merge rights.
C: This is expected, documented default behavior, not a bug.
D: Workflow-approval gating for first-run or untrusted-author PRs is a real, named safeguard — it does not always run automatically.
12Implement guardrails and accountabilityNorthwind Ledger's compliance team is classifying which agent actions require mandatory human sign-off before executing: updating a code comment, versus deleting a customer's stored payment method record via an internal admin tool the agent has access to.Per the guardrails domain, how should these two actions be treated?Tap to reveal the four choices and the correct answer →
This is exactly what the guardrails domain asks for: classify actions by risk to right-size intervention, require explicit authorization for irreversible or compliance-sensitive changes, and avoid approval steps that don't materially reduce risk — treating a comment edit and a customer-data deletion identically would violate both halves of that principle.
A: Uniform approval for every action either over-blocks trivial changes or under-protects a genuinely risky one — the domain explicitly calls for risk-based differentiation instead.
B: An irreversible action touching customer payment data is exactly the kind of compliance-sensitive change the domain says must require explicit authorization.
D: Risk classification is about the action's consequences, not which interface triggered it — an admin-tool deletion is high-risk regardless of whether it touches a code file.
How to use this: if you missed more than 3 of these 12, go back through the domain-by-domain breakdown in the study guide before scheduling — GH-600 leans on operational judgment (what would you configure, approve, or classify), not memorized definitions.
MSCertQuiz sells practice-exam access for GH-600 and other GitHub and Microsoft certifications; this question set is written by the same team that builds those questions.
The Northwind Ledger Story Continues
MSCertQuiz has 500 GH-600 questions covering MCP configuration, memory strategy, evaluation root-causing, multi-agent conflict handling, and guardrail design — weighted to match the real exam.