Save 15% on 2 packs BUY2 · 20% on 3+ BUY3

Skip to content

Exam guide

How to Pass the GitHub GH-600 in 2026: Format, Cost, Domains and Study Plan

Guides › Exam guides

Updated 8 Nov 2026 · 5 min read

On this page
  1. What the GitHub GH-600 is and who it is for
  2. GitHub GH-600 at a glance
  3. What is on the exam
  4. Why people fail it
  5. A study plan that fits the exam
  6. On exam day
  7. Frequently asked questions
  8. Sources

What the GitHub GH-600 is and who it is for

The GH-600, Developing in Agentic AI Systems, is the exam behind the GitHub Certified: Agentic AI Developer certification. It is listed on Microsoft Learn at intermediate level and aimed at AI engineers, developers, DevOps engineers, app makers and solution architects. Microsoft describes the target candidate as someone with subject matter expertise in operating, integrating, supervising and governing AI agents inside production-grade software development lifecycle (SDLC) workflows, with GitHub as the system of record and control plane.

The role covers operating agent workflows inside the SDLC, supervising autonomous behaviour with GitHub controls, evaluating and tuning agent output using scans and artifacts, configuring custom agents and coordinating multi-agent execution safely. Microsoft expects experience with the SDLC, GitHub workflows and controls, code quality, security and review practices, and with coding agents such as GitHub Copilot, MCP servers and agent customisation through custom instructions, custom agents, tools and Copilot setup steps.

GitHub GH-600 at a glance

Item Detail
Exam code GH-600
Time allowed 120 minutes
Passing score 700 out of 1,000, scaled
Exam fee US$99 in the US, priced by the country or region where you sit it
Languages English
Where you sit it Pearson VUE, proctored
Certification valid for 2 years

Microsoft does not publish a question count for the GH-600. It says most of its certification exams typically contain between 40 and 60 questions, but that the number varies by exam. The exam may include interactive components.

What is on the exam

Implement tool use and environment interaction (20 to 25%). The heaviest domain. You select and configure agent tools and their permissions, configure MCP servers including the GitHub remote MCP server, registries and allow lists, and scope an agent to a repository or branch. It also covers invoking an agent from a CI workflow, letting an agent create branches and pull requests, and safe execution with error handling, retries, rollbacks, escalation paths and traceability.

Prepare agent architecture and SDLC processes (15 to 20%). Fitting agents into the SDLC, defining inputs, outputs and success criteria, avoiding common anti-patterns, and keeping planning separate from execution with structured plans that are validated and approved before any action. It also covers setting the degree of autonomy and building in observability and human intervention without slowing delivery.

Perform evaluation, error analysis and tuning (15 to 20%). Defining success criteria and evaluation signals, using automated scanning to generate signals, finding root causes in logs, plans, traces and artifacts, classifying failures such as reasoning errors, tool misuse and context problems, then revising instructions, memory and tool access.

Orchestrate multi-agent coordination (15 to 20%). Applying an orchestration pattern, isolating agents for parallel work, resolving conflicts such as overlapping code changes, and producing artifacts fit for review and audit. It also covers detecting failed or stalled agents, recovery with rollback and human-in-the-loop, and adding, replacing or retiring agents without breaking running workflows.

Manage memory, state and execution (10 to 15%). Choosing between short-term, long-term and external memory, scoping memory to the task, setting expiry and reset rules, persisting state, detecting context drift, resuming work without repeating steps and preventing stale or conflicting context.

Implement guardrails and accountability (10 to 15%). Classifying agent actions by operational, security and compliance risk, assigning autonomy levels, deciding which actions need human judgement, blocking actions that break policy, enforcing least-privilege permissions and keeping approvals that genuinely reduce risk.

Why people fail it

This exam is about operating agents, not just prompting them. Candidates who have used Copilot as a chat assistant but never configured an agent's tool permissions, an MCP allow list or a CI trigger find the heaviest domain unfamiliar, and it carries up to a quarter of the marks.

The second trap is the trade-off built into the guardrails objectives. The blueprint asks for least privilege and human approval, but also for preserving delivery speed by dropping approvals that do not reduce risk. Answers that simply add more oversight can be wrong.

The third is the multi-agent material, which spans conflict handling, recovery, auditability and lifecycle changes. If you have only run one agent at a time, those objectives are reading rather than experience, so set up two agents on the same repository and see what collides.

A study plan that fits the exam

Six weeks suits most developers who already use Copilot or another coding agent. The plan puts tools and environment first because of its weight.

  1. Week 1: agent architecture and the SDLC. Define inputs, outputs and success criteria for one agent task and make it produce a structured plan that must be approved. Take the free GH-600 practice questions to find your gaps.
  2. Week 2: tools and MCP. Add an MCP server to an agent, set tool permissions, try allow lists and registries, and scope an agent to one repository and branch.
  3. Week 3: environments and safe execution. Trigger an agent from a CI workflow, let it open a branch and pull request, then add retries, rollback and an escalation path.
  4. Week 4: memory, evaluation and tuning. Decide what an agent should remember and when it should forget, then diagnose a failed run from its logs and fix the instructions or tool access.
  5. Week 5: multi-agent work and guardrails. Run two agents in parallel, resolve a conflict, then classify actions by risk and set autonomy levels.
  6. Week 6: timed runs. Work through the questions-only copy in the GH-600 practice question pack in 120 minute blocks and read the explanation for every miss.

On exam day

You schedule through Pearson VUE and the exam is proctored. Microsoft recommends registering with a personal Microsoft account, because exam records tied to a work or school account are lost if you leave that organisation. You get 120 minutes and the exam is offered in English only. Microsoft provides an exam sandbox so you can try the question types and interface before you sit it.

Frequently asked questions

What happens if I fail the GH-600?

You can retake it after 24 hours. After that, Microsoft imposes a 14 day wait between subsequent attempts, and you cannot take a given exam more than five times in the 12 months from your first attempt. Check the retake policy on Microsoft Learn before you book.

Is the practice question pack enough on its own?

No. It is practice, not a course. The 58 questions with explanations show how Microsoft frames agent decisions and where your gaps are, but the GH-600 rewards hands-on time configuring agents on GitHub. Build the setups yourself, then use the questions to check yourself and rehearse under time.

How long does the certification last?

Microsoft Learn states that GitHub certifications are valid for 2 years, and that GitHub is moving to Microsoft's recertification process so holders can maintain a certification without retaking the full exam. Check the certification page for the current renewal arrangements.

Try the free questions first, then get all 58 GH-600 practice questions with full explanations for US$39, pass or your money back: get the GH-600 pack.

Sources

The official pages the facts in this guide were checked against.