Nine stages, one project

Agent Gatekeeper tests an AI agent, signs a certificate for it, and revokes that certificate when the agent quietly gets worse. Our demo agent is Harbor, a made-up helpline assistant whose most important answer is a crisis phone number. Every link below opens a real run on a public GitLab page. Settings pages such as approval rules, schedules and feature flags are only visible to members, so we do not link them.

Try the certificate check

Stage	What we do	Open it
1. Plan	One issue per stage on a board, labels for each stage, a milestone for the deadline. Duo turned a vague issue into a plan.	Board, Milestone, Plan issue 43
2. Create	Every change is a merge request that Duo reviews against our written rules. Fourteen custom Duo flows, checked by a test. When the gate failed, Duo's Gate fix flow opened a one-line fix by itself.	Harbor build !41, Fix !43, Flows
3. Verify	Automatic tests on every change, plus the Gatekeeper gate that tests the agent itself. A pretend vendor upgrade changed the crisis number from 555-0142 to 555-0100, and the gate failed it with score 78.	Green main, Red gate job
4. Package	Harbor is built into a container image (a box holding the app and everything it needs) and stored in GitLab's Container Registry.	Package job, Container Registry
5. Secure	GitLab scans our code, leaked passwords, outside libraries, infrastructure files, the image and the running app. A security gate fails the pipeline on a Critical image finding. The image is signed with Sigstore and carries our certificate.	Signing job, Security triage, Triage notes
6. Release	A version tag makes a GitLab release, but only after a person presses play. It links the image and the signed certificate.	Releases
7. Deploy	Staging starts the stored image and checks it answers. Harbor production records a deployment only if Harbor gives the right crisis number.	Production deploy, Environments
8. Monitor	A daily run tests the exact signed production Harbor image again, by its fixed ID (the two runs linked here were made earlier, on the latest image). When a feature flag switched on the pretend vendor model, it revoked the certificate and opened an incident. With the flag off, it passed again.	Revoked run, Incident #37, Restored run
9. Govern	Signed certificates that can be revoked, an Owner sign-off approval rule, a CODEOWNERS file, a manual release step, masked CI variables. Duo audited which written rules GitLab really enforces.	CODEOWNERS, Audit issue 48

What we did not do

No merge in the Harbor run was pressed by a person. Our AI build session merged them from the founder's account, on his instruction. Duo approved the Harbor build.
The daily re-check was triggered by hand for the video. It runs daily at 06:00 UTC, and reaches Harbor through a local port forward because the gate only tests local addresses.
Nobody here is a Maintainer, so we cannot dismiss security findings or use GitLab's built-in observability.
The Owner sign-off approval rule exists but asks for 0 approvals. This project has one member, and an author cannot approve their own change.
The Google Cloud Run deploy job is built but was never run. We have no Google Cloud project.
The live attack test (DAST) is report only and never blocks a merge.
The carbon figure is an estimate from job run times.

Earlier runs on Clerk, a shop assistant, are kept with all their links in the full stages page.

Project: https://gitlab.com/gitlab-ai-hackathon/transcend-october-2026/43380354/agent-gatekeeper. MIT license.