Agentic DevOps on GitHub — Continuous AI PoC

A proof-of-concept that showcases the best of agentic DevOps on GitHub: a real application that ships to Azure through a secure, multi-environment pipeline, with GitHub Continuous AI agents automating triage, documentation, diagnostics, and test improvement.

Continuous AI is to collaborative software what CI/CD is to builds and deployments: routine, automated, AI-assisted work that runs continuously in the background of your repository. This repo demonstrates that vision end-to-end.

---

What this demonstrates

Capability	How it's shown here
Ships to Azure	.NET 10 minimal API → container → Azure Container Apps via Terraform
Secure by default	GitHub OIDC (no cloud secrets), managed identity, ACR without admin creds
Quality gates	dotnet format, xUnit tests, 60% coverage gate, tflint, actionlint, markdownlint
DevSecOps scanning	Trivy (filesystem, IaC config, container image) + CodeQL + Dependabot
Multi-environment	Dev → Test → Prod promotion with manual approval before production
Continuous AI	Four agentic workflows (gh-aw) running on GitHub Copilot

---

Continuous AI agents

These live as Markdown in .github/workflows/ and compile to locked GitHub Actions (*.lock.yml) via gh aw compile. They use safe-outputs — the agent never gets write tokens directly; its proposed actions (comments, draft PRs, issues) are applied by a separate, minimally-scoped job.

Agent	File	Trigger	Continuous AI pillar
Issue Triage	issue-triage.md	Issue opened/reopened	Continuous Triage
Doc Updater	doc-updater.md	Push to main	Continuous Documentation
CI Doctor	ci-doctor.md	CI/CD run fails	Continuous Repair
Test Improver	test-improver.md	Weekly + manual	Continuous Quality

The engine is GitHub Copilot, so no third-party model API keys are required.

Central automation adoption

The agents import shared instructions from DevOpsDerek/workflows. Each source pins its import to a full commit SHA and uses inlined-imports: true, so the compiled lock contains the shared instructions without checking out the catalog at runtime. Local sources retain repository-specific triggers, paths, commands, read permissions, and bounded safe-output configuration; do not copy shared implementations into this repository.

The adopted catalog revision is dac4b81c298cb3ea6821ea312efa5375f42d5ccb.

Issue triage now proposes an existing label and issue type in one comment rather than applying them. Documentation and test improvements produce at most one draft PR per run for human review. CI diagnosis remains failure-only and does not create or apply new labels. No agent may assign issues, change project status, merge, deploy, promote, apply infrastructure, or publish.

The existing meta-lint CI job also calls the central .github/actions/validate-agentic-workflows action using the same immutable catalog revision. It validates and recompiles gh-aw sources and fails on stale locks with only contents: read. A caller-side check also rejects changes to .github/aw/actions-lock.json produced by compilation. Existing .NET build/format/test and 60% coverage gate, test artifacts, Terraform/TFLint, Trivy, CodeQL, workflow/Markdown linting, check names, and triggers remain unchanged. The catalog's narrow checked-script helper is not a replacement for this multi-step CI. CD and environment-bound OIDC promotion remain unchanged; configure the protected production environment and required reviewers as described in setup before deployment.

Use gh-aw v0.89.21, matching the explicit CI validator input:

gh aw compile --validate --actionlint --no-check-update
git diff -- .github/workflows .github/aw/actions-lock.json .gitattributes

Commit each source and generated lock together and review their safe-output tools and permissions, not just compiler success. Never edit locks by hand. Remote imports cached in .github/aw/imports/ are ignored; shared content is embedded in the locks. Update catalog SHAs deliberately after verifying the published paths and contracts, then recompile and review all affected locks.

---

Architecture

flowchart TB
    dev[Developer] -->|push / PR| repo[(GitHub Repository)]

    subgraph GH[GitHub]
        repo --> ci[CI: format, test, coverage, tflint, Trivy, CodeQL]
        repo --> cd[CD: build, scan, promote]
        repo --> agents[Continuous AI agents<br/>triage · docs · ci-doctor · tests]
    end

    cd -->|OIDC, no secrets| azlogin{{Azure OIDC login}}

    subgraph AZ[Azure]
        azlogin --> tf[Terraform: ACR · Log Analytics · Container Apps]
        tf --> aca[Azure Container App<br/>.NET 10 API]
        acr[(Azure Container Registry)] --> aca
        mi[User-assigned Managed Identity<br/>AcrPull] --> aca
    end

    agents -.safe-outputs.-> repo

Loading

Environment promotion

flowchart LR
    build[Build image<br/>+ Trivy scan gate] --> dev[Deploy Dev<br/>auto]
    dev --> test[Deploy Test<br/>auto]
    test -->|manual approval| prod[Deploy Prod<br/>protected]

Loading

Each environment is fully isolated: its own resource group, ACR, Container Apps environment, managed identity, and Terraform state file.

---

Repository layout

.
├── src/Api/                 # .NET 10 minimal API (Task CRUD) + Dockerfile
├── tests/Api.Tests/         # xUnit unit + integration tests
├── infra/                   # Terraform for Azure Container Apps (multi-env)
│   ├── envs/                # dev/test/prod tfvars + backend configs
│   └── bootstrap/           # remote state + GitHub OIDC identities (run once)
├── .github/
│   ├── workflows/           # CI, CD, CodeQL + agentic (*.md → *.lock.yml)
│   └── dependabot.yml
├── AgenticDevOps.sln
└── README.md

---

The application

A small Task API (in-memory store) that is realistic enough to exercise the whole pipeline:

Method	Route	Description
GET	/healthz	Liveness/readiness probe
GET	/tasks	List tasks
GET	/tasks/{id}	Get a task
POST	/tasks	Create a task
PUT	/tasks/{id}	Update a task
DELETE	/tasks/{id}	Delete a task

OpenAPI is exposed at /openapi/v1.json in Development.

---

Run locally

Prerequisites: .NET 10 SDK (see global.json).

dotnet restore AgenticDevOps.sln
dotnet test AgenticDevOps.sln -c Release --settings coverlet.runsettings
dotnet run --project src/Api

On corporate machines with a restricted global NuGet feed, add --configfile nuget.config to the restore command to force nuget.org.

---

Deploy to Azure

Deployment is OIDC-based and runs from GitHub Actions — there are no cloud secrets stored in the repo. See docs/SETUP.md for the full one-time setup (bootstrap state + OIDC identities, GitHub Environments, and variables). In short:

1. terraform -chdir=infra/bootstrap apply — creates the Terraform state account and one GitHub-federated managed identity per environment.
2. Configure GitHub Environments (dev, test, prod) and set the AZURE_CLIENT_ID / AZURE_TENANT_ID / AZURE_SUBSCRIPTION_ID / STATE_STORAGE_ACCOUNT variables. Add required reviewers to prod.
3. Push to main — CD builds, scans, and promotes Dev → Test → Prod.

---

Security posture

• No long-lived cloud credentials — GitHub OIDC federation only.
• No ACR admin user — image pulls use a user-assigned managed identity with AcrPull.
• Defense-in-depth scanning — Trivy (deps, IaC, image) + CodeQL, all reported as SARIF to GitHub code scanning; CRITICAL/HIGH findings fail the build.
• Least-privilege agents — Continuous AI workflows read-only by default; mutations go through scoped safe-outputs jobs.

See SECURITY.md for details.