AgentDeploy¶

An AgentDeploy is the unit of deployment for a single agent. It owns the pods that run the user's agent code together with an injected broker sidecar that handles A2A transport, plus a ClusterIP Service that gives each AgentDeploy a stable DNS name.

AgentDeploys are normally generated by the AgentSet, they are not supposed to be created by the user directly.

Anatomy¶

apiVersion: kynomesh.kyno.sh/v1alpha1
kind: AgentDeploy
metadata:
  name: greeter-planner
spec:
  name: planner
  agentSetName: greeter
  replicas: 2
  scale:
    min: 1
    max: 5
  container:
    image: example/planner:latest
  updateStrategy:
    type: RollingUpdate
    rollingUpdate:
      maxUnavailable: 25%

Pod composition¶

Every AgentDeploy pod is composed of three controller-managed containers plus whatever the user adds:

init-runtime (init container) — writes /var/run/kynomesh/topology.json from the encoded AgentDeploy payload, resolving peer URLs.
agent (Kubernetes-native sidecar with restartPolicy: Always) — runs the user's code. Listens on a Unix Domain Socket the broker connects to.
broker (main container) — terminates A2A traffic and forwards it to the agent over the shared UDS.

Peer discovery¶

The agent discovers peers by reading the topology file:

{
  "pattern": "Supervisor",
  "isEntry": true,
  "peers": [
    {
      "name": "worker1",
      "url": "https://greeter-worker1.ns.svc.cluster.local:8490"
    },
    {
      "name": "worker2",
      "url": "https://greeter-worker2.ns.svc.cluster.local:8490"
    }
  ]
}

For Managed peers, the URL is derived from the sibling AgentDeploy's nomral ClusterIP Service and the broker port.
For External peers, the URL is taken verbatim from the spec.

Rolling updates¶

When the pod spec changes, the controller drives a rolling replacement honoring updateStrategy.rollingUpdate.maxUnavailable. Replica indices are stable across rollouts, so pod DNS names stay predictable.

Kubectl¶

To query AgentDeploy objects with kubectl:

kubectl get agentdeploy # or "ad" as a short name