GitHub Agentic AI Certification (GH-600), the verified reference

The six skill domains, with weights

These are the domain weights as published in the official GH-600 study guide. The wide ranges are Microsoft's, not an estimate. Domain 2 is the heaviest single block and worth front-loading any prep time on.

Weights are intentionally ranges, not exact percentages. The actual exam form draws from each band; Microsoft does not publish a fixed question count for GH-600 because the format may include interactive components.

The largest single domain is about tools, and specifically about MCP

Most coverage of this credential mentions MCP once, in passing, and moves on. The study guide is much more explicit than that. The Domain 2 sub-objective on MCP names four distinct tasks the exam wants you to be able to perform, plus a separate block on tool selection and permission scoping, plus a block on error handling that includes rollbacks and escalation paths. If you flatten that to "know MCP," you have lost the shape of the domain.

The reason MCP weight is high is structural. The certification frames the agent as a client of tools, not as a one-shot text completer. Anything the agent touches outside its own reasoning, whether that is a code search, a CI run, a database, or a third-party API, goes through a tool. MCP is the protocol that defines how those tools attach safely. A GitHub remote MCP server is the canonical hosted shape of that protocol; allow lists and registries are how a platform team keeps agents from reaching the wrong places. The exam treats this as a configuration discipline, not a vague concept.

The SDLC integration shape the exam expects

Domain 1 is the smallest of the six but it is the spine. The study guide repeats two ideas across its sub-objectives: separate planning from execution, and produce inspectable artifacts. Put those together and the agent loop the exam wants you to build looks like this.

Two of those steps, Validate and Review, are the ones most candidates underweight in practice. The study guide is unusually firm here. It explicitly lists "prevent agent action until the agent checked and approved" as a sub-objective, and it lists "configure human intervention for autonomous agents without slowing delivery" as a separate one. That second clause is the harder design problem: gates that do not actually gate are worse than no gates, but gates that pause every action will kill velocity. Domain 6 (guardrails and accountability) is the other place the exam tests whether you can right-size that trade.

Beta-period facts that often get omitted

The honest take on who this is for

The role profile names software developers, platform engineers, DevOps engineers, security engineers, and technical product managers who are already operating agents in real work. The verbs in the audience profile are operating, integrating, supervising, governing. None of those are read-a-book verbs. If you have not set up an MCP server, attached one to a custom agent, scoped its permissions, watched it write a pull request, and then dealt with the cleanup when something went sideways, the exam is going to feel like a vocabulary test you cannot brute-force.

The corollary: this is not a credential for someone who wants their first taste of agentic AI. It is a credential for someone who already runs agents at work and wants the role to be recognized formally. If you are on the "curious about agents" side of the line, the cheapest first step is to actually run one, on a real repo, with a real MCP server, and see what breaks. Everything you learn there maps directly to Domains 2, 4, and 6.

A note on this page

This page is published by mk0r, a small consumer-side example of agentic AI. We make an in-browser app builder where you describe an app in one sentence and watch a Claude agent assemble a working mobile prototype in front of you, no account, no setup. Our agent is way down the spectrum from what GH-600 is testing for: it is one agent, scoped to one task, on one ephemeral sandbox per session. The reason we published the reference above is that we read the study guide closely when designing our own loop, and the version of it on most other pages is too thin to be useful. If you are studying for the exam, the canonical sources linked above are what to use. If you want to feel what a hands-off agent loop is like before you study, mk0r.com is one mouse click and no signup.