Skip to main content

kubeswarm Agent Teams and Pipelines

A kubeswarm SwarmTeam composes multiple AI agents into a workflow on Kubernetes. Three execution modes let you choose the right orchestration pattern for your use case.

Pipeline Mode​

A DAG of steps executed in dependency order. Each step targets an agent.

spec:
  pipeline:
    - role: researcher
      inputs:
        prompt: "Research {{ .input.topic }}"
    - role: writer
      dependsOn: [researcher]
      inputs:
        prompt: "Write about:\n{{ .steps.researcher.output }}"
  output: "{{ .steps.writer.output }}"

Steps pass data via Go template expressions. dependsOn controls ordering.

Dynamic Mode​

Agents delegate to each other at runtime using the built-in delegate() tool.

spec:
  entry: coordinator
  roles:
    - name: coordinator
      canDelegate: [researcher, writer]
    - name: researcher
    - name: writer

The coordinator's LLM decides when to delegate. No fixed DAG.

Routed Mode​

An LLM router picks the best agent from the SwarmRegistry based on the request.

spec:
  routing:
    registryRef:
      name: default
    model: claude-haiku-4-5-20251001
    fallback: general-agent

No pipeline, no roles - the team adapts to what each request needs.

Execution Records​

Every run creates a SwarmRun - an immutable execution record with inputs, outputs, token counts and phase transitions. SwarmRun is to SwarmTeam what Job is to CronJob.

kubectl get swrun -w
kubectl describe swrun my-run-20260403