How should I prepare for GH-600?

Measure yourself against the six domains first, then study in weight order starting with tool use and environment interaction (20-25%). Use Microsoft's three GH-600 Learn modules and GitHub's documentation, and practise in a test repository: scope an agent, give it restricted tools and an MCP server, trigger it from CI, and add guardrails.

Is there an official GH-600 practice test?

Not yet. Microsoft offers a free practice assessment for GH-300 but none for GH-600. A readiness test weighted by the published GH-600 domains is the closest substitute.

Which GH-600 domain should I study first?

Implement tool use and environment interaction, at 20-25% the heaviest. It covers tool permissions, MCP servers and allow lists, repository and branch scoping, CI-invoked agents, and error handling, retries, rollbacks and traceability.

Know if you're actually ready. Take the GitHub GH-600 quiz → get your AI readiness report.Take the free test →

Study guide · GitHub GH-600

How to Prepare for GH-600: A Study Plan in Weight Order

Six domains, ordered by weight, each with the official material Microsoft points to and one hands-on task that covers most of it.

Examifyr·2026·7 min read

Start by measuring, not reading

GH-600 spreads its marks across six domains, and most people preparing for it are already strong in one or two. Take a domain-weighted readiness test first so the plan below is ordered by your gaps, not just by the blueprint. Then work through the domains from heaviest to lightest, and spend most of the time in a test repository rather than in documentation.

Official material

Microsoft's study guide points to three Learn modules and to GitHub's documentation for each domain. Use these as the source of truth: the exam may include Preview features if they are commonly used, so recent GitHub docs beat older third-party courses.

Microsoft Learn modules
  - Foundations of Agentic AI in GitHub
  - Designing Agent Architecture and SDLC Integration
  - Tooling, MCP, and Agent Execution Environments

GitHub Docs (per domain): custom agents, Copilot memory,
cloud agent guardrails and risks, organisation setup for custom agents

1. Tool use and environment interaction (20-25%)

The heaviest domain. In a test repository: give an agent a restricted tool list, add an MCP server, configure an MCP allow list, scope the agent to one repository and to a branch, trigger it from a CI workflow, and let it open a pull request. Then break something on purpose and practise the other half of the domain: error handling, retries, rollbacks, escalation, and tracing which agent did what.

2. The three 15-20% domains

Agent architecture and SDLC: make an agent produce a structured plan, validate it, and block action until it is approved; define inputs, outputs and success criteria. Evaluation, error analysis and tuning: decide what a successful agent task looks like, use automated scanning to generate evaluation signals, classify failures as reasoning errors, tool misuse or context and environment problems, then tune instructions, memory or tool access. Multi-agent coordination: run agents in parallel with isolation, resolve overlapping changes and contradictory outputs, and add, replace or retire an agent without breaking an active workflow.

3. The two 10-15% domains

Memory, state and execution: choose between short-term, long-term and external memory, capture progress as durable artifacts so an agent can resume without repeating itself, and detect drift on long runs. Guardrails and accountability: classify agent actions by operational, security and compliance risk, require explicit approval only for irreversible or compliance-sensitive changes, enforce least privilege, and avoid approvals that slow delivery without reducing risk.

NoteThe guardrails domain rewards balance, not caution. "Require human approval for everything" is usually the wrong answer; the blueprint explicitly asks you to preserve velocity by minimising approvals that do not materially reduce risk.

When to book

Book when a readiness measurement puts you near the line in every domain, not when you finish a course. GH-600 has no official practice assessment, so a blueprint-weighted test is the closest thing to a dress rehearsal. Microsoft allows a retake 24 hours after a first failure, with longer waits after that, so the first attempt is the one to time well.

Exam tipKeep one test repository for the whole of your preparation and build every hands-on task in it. By the end it holds a scoped, CI-triggered agent with an MCP allow list, a plan-approval gate, an evaluation step and a rollback path, which is close to the whole blueprint in one place.