The most interesting new exam of the year for working engineers. It assumes agents already write your code, and asks whether you can run them safely: scopes, guardrails, evaluation and recovery. Very little of it is GitHub trivia.
Take it if
Your team runs coding agents against real repositories.
Skip it if
You have not yet used a coding agent on a shared codebase. Start with GH-300.
Before you start
Check yourself against this list first. If something is missing, close that gap before you pay for an exam.
You have used a coding agent on a shared repoStart with GH-300 if you have not.
You understand GitHub workflows and controlsBranch protection, Actions and reviews.
You have used MCP serversAdd one to an agent first. Learn more
SDLC, GitHub workflows, coding agents, MCP servers and agent customization
If you fail
Wait 24 hours after a first attempt
What the exam covers
Prepare agent architecture and SDLC processes15–20%
Implement tool use and environment interaction20–25%
Manage memory, state, and execution10–15%
Perform evaluation, error analysis, and tuning15–20%
Orchestrate multi-agent coordination15–20%
Implement guardrails and accountability10–15%
What to focus on
Keeping planning apart from execution, and blocking action until a plan has been checked.
Tool work, the largest domain: permissions, MCP registries and allow lists, repository and branch scope.
Memory: short-term, long-term and external, with expiry and reset rules.
Finding root causes in logs, plans and traces, and telling reasoning errors from tool misuse.
Multi-agent isolation, conflict detection, and recovery with rollback and human steps.
Sorting actions by risk so people review what matters, without slowing delivery on what does not.
Everything on one mapEvery topic and key idea in this certification.
mindmap
root((GH-600))
Agents in the SDLC
Structured plans
Success criteria
Inspectable artifacts
Tools and environments
Least access
MCP allow lists
Failure handling
Memory and state
Memory types
Expiry
Drift
Evaluation and tuning
Signals
Root causes
Tuning
Multi-agent work and guardrails
Isolation
Risk-based approval
Recovery
A safe path for agent workGitHub says a cloud agent pushes to a single branch, and a human must review and merge its draft pull requests.
sequenceDiagram
participant D as Developer
participant A as Cloud agent
participant R as Repository
participant S as Security checks
D->>A: Assign the task
A->>R: Push to its own branch
A->>S: CodeQL and secret scanning
S-->>A: Findings
A->>R: Open a draft pull request
D->>R: Review and merge
Lesson 1 of 5, 15–20% of the exam
Agents in the SDLC
Separating planning from action is the core habit here.
Structured plans
The agent outputs a plan you can check before it acts.
Success criteria
Define inputs, outputs and what done means.
Inspectable artifacts
Agent work should show up in normal tools, such as pull requests.
Plan before actionThe agent cannot act until the plan is checked.
stateDiagram-v2
direction LR
[*] --> Plan
Plan --> Review
Review --> Plan: Rejected
Review --> Execute: Approved
Execute --> Artifacts
Artifacts --> [*]
Common trap
Letting an agent edit before anyone has seen its approach.
Try it yourself
Have an agent produce a plan, then block action until it is approved.
The largest domain, and most of it is permissions.
Least access
Give each agent only the tools and scope it needs.
MCP allow lists
Restrict which servers can be used.
Failure handling
Build in retries, rollbacks and escalation paths.
Layers that limit an agentEach one holds even if the prompt fails.
flowchart TD
A["Agent"] --> B["Allowed tools"]
B --> C["MCP allow list"]
C --> D["Repository scope"]
D --> E["Single working branch"]
E --> F["Branch protection on main"]
F --> G["Human merges"]
Common trap
Trusting a prompt to keep an agent away from main. Use branch protection.
Try it yourself
Scope an agent to one repo and one branch, then try to break out.
Choose between short-term, long-term and external.
Expiry
GitHub says Copilot Memory facts expire by default after 90 days.
Drift
Save decisions as artifacts, so work resumes without repeating itself.
Agent memory optionsChoose by how long the fact must survive.
flowchart LR
A["Short-term: this session"] --> B["Long-term: across sessions"]
B --> C["External: shared artifacts"]
A -.-> D["Lost when the session ends"]
B -.-> E["Needs expiry and pruning"]
C -.-> F["Survives across tools"]
Common trap
Assuming memories last forever or are always private. Repo facts are shared with collaborators.
Try it yourself
Have an agent write its progress to a file, then resume in a new session.
Each parallel agent gets its own branch or workspace.
Risk-based approval
Require people for irreversible changes, and drop approvals that do not reduce risk.
Recovery
Detect stalls, then roll back or escalate.
Sort actions by riskKeep people on the decisions that matter.
flowchart TD
A{"Action risk"} -->|"Low, reversible"| B["Let the agent proceed"]
A -->|"Medium"| C["Review in the pull request"]
A -->|"High or irreversible"| D["Explicit human approval"]
A -->|"Violates policy"| E["Block"]
Common trap
Choosing "add more approvals" every time. The blueprint values delivery speed too.
Try it yourself
Run two agents in parallel, then sort their actions by risk.
Finish Foundations of Agentic AI in GitHub, Designing Agent Architecture and SDLC Integration, and Tooling, MCP, and Agent Execution Environments.
Week 2: configure agents
Write a custom agent with instructions and restricted tools. Add an MCP server with an allow list.
Week 3: run them for real
Invoke an agent in a CI workflow, scoped to a branch. Have it open pull requests, then build in an approval step.
Week 4: failure and multi-agent
Force agent failures and classify each root cause from the logs. Run two agents in parallel and resolve a conflict.
Exam week
Take the mock below and review each guardrail decision you missed.
Exam day and after
Microsoft Learn is not available during GitHub exams.
Expect questions where more approvals is the wrong answer. The blueprint explicitly asks you to preserve delivery speed by removing approvals that do not reduce risk.
Test yourself8 practice questions, scored by exam topic.
Exam facts checked on October 31, 2026 against GH-600 study guide. Lessons are our own explanations, so use the official docs as the final word, and confirm price and dates before booking.
