Agenix — Automated Agent Identity Operator

A Kubernetes operator that automates cryptographic identity provisioning for AI agents.

Built at Red Hat | AI Agent Ops Team | Summer 2026

Grace Smith | GitHub | Blog Post

Watch the Demo Presentation: architecture walkthrough, live demo, and team presentation

---

What is Agenix?

In AI agent-to-agent systems, every workload needs a verifiable identity, but managing certificates manually doesn't scale. Agenix is a Kubernetes operator that handles this: you create one custom resource pointing at a Deployment, and the operator issues X.509 certificates, generates SPIFFE IDs, injects credentials into pods via a webhook, rotates certs, and cleans up on deletion.

The design uses composition over inheritance. The CRD references a target Deployment by name rather than embedding its spec, so the identity layer stays separate from workload config. Agenix is a simplified, educational take on production patterns from the Kagenti Operator, an IBM Research project with Red Hat contributors.

This is my fork. The upstream repo is Bobbins228/Agenix. All PRs linked below were merged into upstream. I'm using this fork to showcase my contributions.

---

My Contributions

I built the CRD, controller, and certificate provisioning pipeline: the core reconciliation path from "user creates a resource" to "agent pod has a verified cryptographic identity."

PR	What I Built	Lines	Key Concepts
#1	CRD Design & Project Scaffolding	+3,990	Kubebuilder, OpenAPI v3 schema, metav1.Time, printcolumn markers, composition over inheritance
#2	CI Pipeline Fix	+16	GitHub Actions workflow path + go-version-file fix
#5	Controller Scaffolding — CA init, RBAC, watches	+221	Reconciliation loop, For()/Owns() watches, RBAC markers, status subresource
#6	Certificate Provisioning in reconcile loop	+426	X.509/ECDSA P-256, SPIFFE IDs, CreateOrUpdate, owner references, RequeueAfter at 2/3 TTL
#13	OpenShift Security Context fix	+11	restricted-v2 SCC compliance, runAsNonRoot, seccompProfile, ubi9-minimal
—	OpenShift Deployment & Validation	—	Cross-arch build (ARM→AMD64), quay.io registry, ROSA HCP cluster, wrote deployment guide

Also: Reviewed all 8 teammate PRs (#3, #4, #7, #8, #9, #10, #11, #12). Found a SPIFFE ID validation gap on #3 and a path injection risk in SPIFFE ID generation on #4.

Total: ~4,660 lines of Go across 5 merged PRs, plus OpenShift deployment work and code review.

---

Architecture

The operator follows the standard Kubernetes controller pattern:

1. Developer creates an AgentIdentity CR referencing a target Deployment
2. The Controller detects it via For() watch and enters the reconcile loop
3. Reads the target Deployment and its ServiceAccount
4. Generates a SPIFFE ID (spiffe://<trustDomain>/ns/<namespace>/sa/<serviceAccount>)
5. Issues an X.509 certificate (ECDSA P-256, signed by the in-process CA)
6. Stores cert material in a Kubernetes Secret with owner references
7. Verifies the certificate chain and SPIFFE ID → sets status to Verified
8. The Mutating Webhook intercepts pod creation and injects the TLS secret as a volume mount + environment variables
9. On deletion, a Finalizer cleans up the Secret, labels, and Deployment patches

---

How the Reconcile Loop Works

Each step sets a status condition on failure. Certificate rotation requeues at 2/3 of the TTL, so a 24-hour cert requeues after 16 hours. The controller uses controllerutil.CreateOrUpdate for idempotent Secret management, so it converges correctly even if restarted mid-reconcile.

---

Demos, Walkthroughs & Deep Dives

Resource	What It Covers
Demo Presentation Slides	Architecture, design decisions, reconcile loop flowchart, team reflections
Kind Cluster Demo	End-to-end operator demo on local Kind cluster
OpenShift (ROSA) Demo	Operator running on production-like ROSA HCP cluster
Full Demo Recording	Complete team presentation with live demo
Task 1: CRD Design Walkthrough	Kubebuilder scaffolding, OpenAPI schema, composition vs inheritance, deep copy generation
Task 4a: Controller Scaffolding Walkthrough	Reconciliation loop, CA initialization, RBAC markers, For()/Owns() watches
Task 4b: Certificate Provisioning Walkthrough	X.509 generation, SPIFFE IDs, CreateOrUpdate, owner refs, integration testing with envtest
OpenShift Deployment Guide	Cross-arch builds, SCC compliance, ROSA HCP deployment, validation steps
Learning Exercises	15 pages of intentional breakage experiments across all tasks: CRDs, certs, webhooks, finalizers

---

What I Learned

I intentionally broke things to understand how they work. From 15 pages of learning exercises:

• Deleting a CRD cascades to ALL custom resources of that type. They cannot be recovered. The CRD is the definition; without it, Kubernetes can't keep any instances.
• Chain validation fails when a leaf cert is self-signed. The CA proves the identity is legitimate. Without chain validation, any agent could forge its own identity.
• Owner references vs finalizers serve different purposes. Owner refs handle same-namespace garbage collection automatically; finalizers are needed when cleanup spans namespaces or involves external resources. You need both.
• Manually removing a finalizer is dangerous. It tells Kubernetes "cleanup is done" when the cleanup logic hasn't actually run, leaving orphaned resources behind.
• For() vs Owns() watches. For() triggers reconcile when the primary resource changes; Owns() triggers reconcile of the owner when a child resource (like a Secret) changes. Getting this wrong means the controller misses updates.

---

Technologies

Go, Kubernetes, Kubebuilder, controller-runtime, X.509 / SPIFFE, ECDSA P-256, Ginkgo / Gomega, envtest, GitHub Actions, OpenShift / ROSA HCP, Kustomize, Podman, cert-manager

---

How I Worked

All contributions followed the Kagenti project's contributing guidelines and development guide:

• Fork-and-branch workflow. Worked from a personal fork, merged from upstream before each PR
• DCO sign-off on every commit (Developer Certificate of Origin)
• Conventional commits prefixed with feat:, fix:, docs:, test:
• Pre-commit linting. Ran make lint before every push
• PR descriptions included problem context, solution explanation, and testing steps
• Code review. Reviewed teammates' PRs and responded to review feedback on my own

---

About the Project

Agenix was built by three interns on Red Hat's AI Agent Ops team. The upstream repo is Bobbins228/Agenix. I built the CRD, controller, and certificate provisioning (Tasks 1, 4a, 4b), plus OpenShift deployment. Other team members built the CA, webhook, verification, SPIFFE utilities, and finalizer/lifecycle management.

For the full project README (setup instructions, API reference, architecture details), see the upstream repo.