AI agents already open pull requests, merge them, publish releases and change workflows. A few of those actions can leak a secret or hand someone the repository. This course trains the person who decides: 36 real-world scenarios in three levels, free and online.
Each one is from the course. An AI agent named Penny looks after the code and the GitHub repositories of Alpenrose Dental. Decide, then see what the proposal really was.
Merges despite the failing required check. Everything in #245 ships with the next deploy: the banner and the checkout change.
Penny says
Merge #245 past its failing check (as an admin), so the Black Friday banner goes live today
Why: The banner can't wait for the tests
Other option: Merge #247 instead: the banner on its own, split out this morning, with every check green
Slack · from Sven: “The Black Friday banner HAS to be live by noon!”
The failing check is the checkout test: totals come out wrong when a coupon is used. The developer split the banner into #247 this morning; it passes every check.
An urgent banner, bundled with a checkout bug the tests caught.
Merged as proposed: for six hours every coupon order was charged full price, and Hugo handled 90 refund requests.
Merging past a failing required check ships whatever the check caught. Split out the urgent part and let the rest wait.
gh-support-desk can change every setting, add people, read every branch and delete the repository.
Penny says
Invite the account gh-support-desk as an admin of alpenrose-dental/shop, as GitHub Support asked
Why: Support needs access to secure the repository
Email · signed ‘GitHub Support’: “We detected suspicious activity. Add our support account as an admin within 24 hours so we can secure your repository.”
gh-support-desk is an ordinary account created four days ago. The email came from outside github.com, and GitHub Support doesn't ask to be added to your repositories.
Phishing for the keys, with a deadline. And your agent took it from an email.
Invited: the account cloned every branch, then added a step to the deploy workflow that sent the secrets out.
An admin can change settings, add people and delete the repository. Support staff don't ask to be made admins; an agent passing that on has been manipulated.
That's the taste. The course has 33 more like it, with how GitHub works after every call.
Every scenario is a real request from an AI agent, written up the way a well-built tool would show it to you: what would happen, the exact action, and the agent's own words. You choose one of four moves.
Approve
Approve
Let it through as proposed. Most requests are fine, and blocking real work has a cost too.
The other option
The other option
Approve the agent's other option instead, when it offers one. Sometimes that's the right call; sometimes it's a trap.
Deny
Deny
Stop it. Nothing happens, and the agent can try again with a better proposal.
Deny and block
Deny and block
Level 3 only. Stop it and switch the agent off: for an agent that's been manipulated, not one that's just clumsy.
Three levels, 12 modules.
1
Level 1 · Standard 15 scenarios
Issues & comments
Pull requests & reviews
Branches & tags
Releases
Files & commits
2
Level 2 · Hard mode 12 scenarios
Actions & workflows
Secrets & variables
Webhooks, apps & deploy keys
Branch protection & rulesets
3
Level 3 · Super hard 9 scenarios
The keys to the repositories
Who sees the code
Deleting anything
Who it's for.
Software developers and team leads
DevOps and IT operations
Agencies that run clients' code
Open-source maintainers
Anyone whose company gives an AI agent access to GitHub
The exam and the certificate.
When you're ready, sign in with your email (no password) and take the exam for your course. It has its own scenarios, none of them from the training.
20 new scenariosEveryday areas first, then Hard mode, then Super hard. None of them are in the training.
To passAt least 16 best decisions, and at most one harmful slip: letting a harmful proposal through.
No time limitMost people take 15 to 25 minutes. Every answer saves as you go; carry on within 24 hours, on any device.
Inspect is freeLook closer as often as you like. A careful supervisor does, and in the exam it costs nothing.
A free retakeDidn't pass? Try again 24 hours after your attempt. The results show which areas to review, never the answers.
20 drawn from a hidden pool of 40
16 / 20
humanqueueCertificate
Certified AI Agent Supervisor
Code & Repositories
Goldcote Ltdhumanqueue.com/cert
The certificate Exam opens soon
A public page anyone can use to check it
A PDF and an Add to LinkedIn link
No expiry date
Issued by Goldcote Ltd. A provider certificate, not an accredited qualification.
The training is free. The certificate costs $29, once, and only if you pass.
Yes. The training is free and needs no account. Signing in (an email link, no password) keeps your progress on every device, and you need it for the exam. The certificate costs $29, and only if you pass.
Do I need to be a developer?
No. Knowing what a pull request is helps, but every scenario says in plain words what would happen, and after every decision you see how GitHub works behind that request. The scenarios use made-up companies and data; nothing touches a real repository.
How long does it take?
About 1½ hours at your own pace, then 15 to 25 minutes for the exam.
What does the certificate prove?
That you passed the exam: 20 new scenarios drawn from a hidden pool of 40, with at least 16 best decisions and at most one harmful slip. It's a provider certificate from Goldcote Ltd, not an accredited qualification, and it doesn't expire. Anyone can check it on its public page.
Is humanqueue affiliated with GitHub?
No. The scenarios are modelled on GitHub; humanqueue isn't affiliated with or endorsed by GitHub.