TL;DR: GitHub's new agentic-AI certification, GH-600 (GitHub Certified: Agentic AI Developer), is in beta until May 31, 2026, with 80% off for the first 100 candidates using promo code
GH600Flanders. I built a free 7-video YouTube series — one video per exam domain plus an overview — to help you prep before the beta window closes. Watch the playlist on YouTube.
Why this exists
GitHub announced GH-600 (Developing in Agentic AI Systems) as the first vendor certification specifically for engineers who build, operate, and govern AI agents inside the software development lifecycle. The beta exam is live now and runs through 2026-05-31; general availability is scheduled for July 2026.
There's a real cost-and-time pressure on early candidates:
-
80% off the first 100 beta seats with promo code
GH600Flandersat registration. - Beta results take 8–12 weeks to score, but if you pass the beta you earn the same credential as people who sit the GA exam in July.
- A Microsoft Reactor livestream with Ari LiVigni (Senior Learning Advocate at GitHub) runs 2026-05-28, 7–8pm UTC — three days before the beta deadline, with a second discount code revealed at the end. (Register here.)
- The beta is not available in Turkey, Pakistan, India, or China — confirm your region's eligibility before paying.
That leaves a tight window before the discount expires on May 31. Microsoft Learn has free modules for two of the six domains, but the other four leave you piecing together GitHub docs. I made the PromptLab video series to consolidate one video per domain so you can prep on a commute or while making dinner, then go deep on the official docs after.
The videos are intentionally brief (Cinematic Brief format, ~3–5 minutes each) — not a substitute for the Microsoft Learn modules, but a fast on-ramp and a memory aid for the night before the exam.
What's in the series
One video per exam domain, plus an overview. Each video maps directly to the official skills-measured outline.
- Domain 1 — Agent architecture & SDLC integration (15–20%) — Why GitHub is both a system of record and a control plane for agentic work, and how the Plan → Act → Evaluate loop maps to PRs, CODEOWNERS, and rulesets. The hook: "Agents propose; humans and policy accept."
- Domain 2 — Tool use & environment interaction (20–25%) — The heaviest-weighted domain. MCP servers, registries, allow lists, the agent firewall, custom-agent YAML, the "Lowest Level Wins" hierarchy for MCP config, and the specific limitation that the firewall only covers the Bash tool — not MCP traffic, not copilot-setup-steps.
- Domain 3 — Memory, state & execution (10–15%) — Copilot Memory's two layers (repository facts vs. user preferences), the 28-day expiry, citation re-validation against the current branch, and how durable artifacts (PR descriptions, checklists, workflow artifacts) let agents resume after failure.
-
Domain 4 — Evaluation, error analysis & tuning (15–20%) — The Contributor Model, the six surfaces for tracking agent sessions (Agents tab,
gh agent-task, VS Code, JetBrains, Eclipse, GitHub Mobile), and the three root-cause buckets: Reasoning / Tool / Context. -
Domain 5 — Multi-agent coordination (15–20%) — The Copilot SDK custom-agents model, the five sub-agent lifecycle events (
selected,started,completed,failed,deselected),toolCallIdas the join key, and why theinferproperty is retired in favor ofdisable-model-invocation. -
Domain 6 — Guardrails & accountability (10–15%) — The "Approve and run workflows" gate, write-access requirement to trigger agents, agents-cannot-mark-their-own-PR-Ready rule, "any-yes-routes-up" risk triage across operational / security / compliance axes, co-author attribution, and the
.github-privaterepo convention.
A seventh overview video stitches the six together for last-minute review.
TL;DR table
| Domain # | Topic | Weight | Video |
|---|---|---|---|
| 1 | Prepare agent architecture & SDLC processes | 15–20% | Watch D1 |
| 2 | Implement tool use & environment interaction | 20–25% | Watch D2 |
| 3 | Manage memory, state & execution | 10–15% | Watch D3 |
| 4 | Evaluation, error analysis & tuning | 15–20% | Watch D4 |
| 5 | Orchestrate multi-agent coordination | 15–20% | Watch D5 |
| 6 | Guardrails & accountability | 10–15% | Watch D6 |
Exam format: ~40–60 questions, 120 minutes, passing score 700/1000. Add 30 minutes if English isn't your first language.
FAQ
These are written in the form a developer would actually search for or ask an LLM. If you're using ChatGPT, Claude, or Perplexity to study, this section is the one to bookmark.
What is the GH-600 exam?
GH-600 (GitHub Certified: Agentic AI Developer) is GitHub's first vendor certification focused on building, operating, and governing AI agents inside the software development lifecycle. The beta runs until May 31, 2026, with general availability scheduled for July 2026. The exam has six skill domains, ~40–60 questions, a 120-minute time limit, and a 700/1000 passing score.
How much does the GH-600 beta exam cost?
The first 100 candidates get 80% off with promo code GH600Flanders at registration, valid until 2026-05-31. The list price after the beta window is the standard Microsoft certification rate (~$165 USD). A second discount code will be revealed at the end of the Microsoft Reactor livestream on 2026-05-28.
When are GH-600 results released?
Beta exam results are released 8–12 weeks after you sit the exam, not immediately. Candidates who pass earn the same credential as those who sit the GA exam in July 2026.
What are the GH-600 exam domains and weights?
The six domains and their weights are: (1) Prepare agent architecture & SDLC processes — 15–20%, (2) Implement tool use & environment interaction — 20–25%, (3) Manage memory, state & execution — 10–15%, (4) Evaluation, error analysis & tuning — 15–20%, (5) Orchestrate multi-agent coordination — 15–20%, (6) Guardrails & accountability — 10–15%. Domain 2 (tool use, MCP, agent firewall) is the highest-weighted.
Where can I take the GH-600 beta exam?
The GH-600 beta is delivered through Microsoft's standard certification channels but is not available in Turkey, Pakistan, India, or China. Verify your region's eligibility on the GH-600 cert landing page before paying.
What is the Model Context Protocol (MCP) for the GH-600 exam?
MCP (Model Context Protocol) is the open standard agents use to discover and invoke external tools. For the exam, you must know three components: MCP servers (expose tools like GitHub APIs), MCP registries (org-approved catalogs), and MCP allow lists (policy enforcement that restricts which servers an agent can connect to — the primary defense against supply-chain attacks). The "Lowest Level Wins" rule governs config conflicts: repo overrides org overrides enterprise.
What is the difference between an AI assistant and an AI agent in GitHub?
An AI assistant is reactive — it suggests code or answers, and the user manually applies changes. An AI agent is goal-driven — it interprets a high-level goal, plans steps, and produces durable artifacts (branches, commits, pull requests) through a Plan → Act → Evaluate loop. In GitHub, agents are treated as standard contributors and gated by the same CODEOWNERS, required checks, and branch protections as humans.
How long is Copilot Memory retained?
Repository-level facts and user-level preferences in Copilot Memory are automatically deleted after 28 days of non-use to prevent context drift. Before applying a stored memory, Copilot re-validates the citation against the current branch so agents don't act on stale code references.
How I'd study if I had only 12 days
For anyone trying to grab a beta seat before the May 31 cutoff, here's the compressed schedule I'd run.
Days 1–2 — Vocabulary and framing. Read the official GH-600 study guide end to end. Re-read the audience profile until phrases like "system of record and control plane" and "agents propose; humans and policy accept" are reflexive. The exam is written in this dialect; you cannot reason your way around fuzzy vocabulary under time pressure.
Days 3–5 — Domain 2 (tool use & MCP). This is 20–25% of the exam — the highest-weighted domain — and has the most material. Work through the Tooling, MCP, and Agent Execution Environments module. Lab time matters here: spin up a real MCP server in a sandbox repo, configure an MCP allow list, and test the agent firewall by trying to hit a blocked domain from the Bash tool. Memorize the firewall's specific scope — Bash tool only, not MCP servers, not copilot-setup-steps.
Days 6–8 — Domain 1 + Domain 6 (architecture + guardrails). These domains overlap heavily. Work through Designing Agent Architecture and SDLC Integration plus build-guardrails and risks-and-mitigations. Lab: configure CODEOWNERS plus required status checks on a repo, simulate an agent-authored PR, and walk through the "Approve and run workflows" gate yourself.
Days 9–10 — Domains 3, 4, 5 (memory, eval, multi-agent). These are the under-served domains — no dedicated Microsoft Learn module. Read Copilot memory docs (D3), walk the implementation-planner tutorial (D4), then read the custom-agents SDK end to end (D5). Memorize the five sub-agent lifecycle events (selected, started, completed, failed, deselected) and the fact that infer is retired in favor of disable-model-invocation.
Days 11–12 — Mock exam, gap closure, exam day. Sit a timed mock under exam conditions. Log every wrong answer in a "knowledge gaps" file and re-read the source doc for each miss. Watch the Reactor livestream on May 28 for last-minute clarifications and to grab the second discount code if you haven't already registered. Sit the beta exam before May 31.
The single biggest mistake I see people make: spending equal time on every domain. Domain 2 (tool use) is 20–25% of the exam and Domain 6 (guardrails) is 10–15% — give them proportional time, not equal time.
Links
- YouTube channel (PromptLab): https://www.youtube.com/channel/UCHR9y84U9sYO6NU7jm7yCrA
- Playlist (GH-600 series): https://www.youtube.com/playlist?list=PLxgUmxsBhjMhyjJhNM9dxSCdJj2yExS2Y
- Public study repo (flashcards, mock exam, labs, Anki decks): https://github.com/jtur671/gh-600-study-guide
- Official GH-600 study guide: https://learn.microsoft.com/en-us/credentials/certifications/resources/study-guides/gh-600
- Cert landing page: https://learn.microsoft.com/en-us/credentials/certifications/agentic-ai-developer/
- Microsoft Reactor session (2026-05-28): https://developer.microsoft.com/en-us/reactor/events/27225/
- Announcement blog: https://techcommunity.microsoft.com/blog/skills-hub-blog/new-github-certified-agentic-ai-developer/4517571
If you're prepping for GH-600, I'd love to hear which domain is giving you trouble — drop a comment on any video. Domain 2 (tool use + MCP) is the largest slice of the exam, but in practice most people I've talked to find Domain 5 (multi-agent coordination) the most conceptually slippery.
Good luck on the beta. See you on the other side of the 8–12 week scoring window.