Skip to content

README.md

Latest commit

 

History

History
112 lines (91 loc) · 4.87 KB

README.md

File metadata and controls

112 lines (91 loc) · 4.87 KB
Raw

deploy/

Raw, per-component Kubernetes manifests for agentctl plus a local end-to-end walkthrough. Use this directory for development, Kustomize overlays, and reading each object in isolation.

For a production install, use the Helm chart: charts/agentctl. One helm install brings up every control-plane component and issues all TLS (the aggregated APIServer serving cert, the admission webhook cert, each agent's serving identity, and the gateways' certs) through cert-manager with automatic caBundle injection and rotation.

Layout

deploy/
  crds/                    # generated CRDs — regenerate with `cargo run -p agentctl-crdgen`
    agent.yaml
    agentfleet.yaml
    modelpool.yaml
    mcpserverset.yaml
  operator/                # in-cluster operator install (namespace + RBAC + Deployment)
    namespace.yaml
    rbac.yaml              # ServiceAccount + least-privilege ClusterRole(+Binding)
    deployment.yaml
    Dockerfile             # distroless runtime over the host-built release binary
    kustomization.yaml
  apiserver/               # aggregated APIServer (management verbs): Deployment + RBAC + APIService
  admission/               # validating + mutating webhooks: Deployment + RBAC + webhook config
  gateway/                 # A2A gateway: Deployment + RBAC + signing Secret
  modelgateway/            # intelligence broker: Deployment + RBAC
  mcpgateway/              # tools broker (Dockerfile)
  coordination/            # work-distribution MCP server (Dockerfile)
  scaler/                  # KEDA external scaler (Dockerfile)
  postgres/                # bundled durable store for gateway/modelgateway/coordination
  hardening/               # NetworkPolicies (default-deny + control-plane allow) + mTLS helper
  examples/                # sample CRs: Agent, AgentFleet, ModelPool, and mock fixtures

The crds/ files are generated from the Rust CRD types in the agent-api crate. Never hand-edit them — run cargo run -p agentctl-crdgen to regenerate, and CI fails on drift.

In-cluster operator install

Run the operator inside the cluster with its own ServiceAccount and RBAC:

# build + load the operator image into kind
cargo build --release -p agentctl-operator --bin agentctl-operator
docker build -f deploy/operator/Dockerfile -t agentctl/operator:dev .
kind load docker-image agentctl/operator:dev --name agentctl

# install CRDs + operator
kubectl apply -f deploy/crds/
kubectl apply -k deploy/operator/
kubectl -n agentctl-system rollout status deploy/agentctl-operator

# use it
kubectl apply -f deploy/examples/agent-once.yaml
kubectl get agents
kubectl logs -n agentctl-system deploy/agentctl-operator

The deploy/operator/ overlay is the only self-contained Kustomize install here. The other components (apiserver/, admission/, gateway/, modelgateway/, postgres/, hardening/) carry raw manifests you can apply individually, but they depend on cert-manager-issued TLS Secrets and shared configuration that the Helm chart wires for you. For a complete control plane, install the chart.

Local end-to-end (kind)

The fastest development loop runs the operator out of cluster against a kind cluster, reading your kubeconfig:

# 1. cluster + CRDs
kind create cluster --name agentctl
cargo run -p agentctl-crdgen          # (re)generate deploy/crds/
kubectl apply -f deploy/crds/

# 2. run the operator (reads your kubeconfig)
cargo run -p agentctl-operator

# 3. create an Agent and watch it reconcile
kubectl apply -f deploy/examples/agent-once.yaml
kubectl get agent demo -o jsonpath='{.status}'   # Ready=True / phase=Ready
kubectl get jobs                                  # the operator rendered a Job
cargo run -p agentctl-cli -- get                  # NAME MODE READY PHASE AGE
cargo run -p agentctl-cli -- describe demo

# 4. deletion → finalizer + owner-ref GC reclaim the Job
kubectl delete agent demo
kubectl get jobs                                  # gone

# teardown
kind delete cluster --name agentctl

Note on the example image. agent-once.yaml uses busybox as a placeholder. There is no conformant-agent image in this repo, so the rendered Job's pod will fail (busybox does not understand the agent arguments). That is expected: the walkthrough exercises the control plane — reconcile, render, apply, status, and garbage collection — not a running agent. Point .spec.image at a real conformant agent (the reference implementation is agentd) to run a live agent.

Related documentation