Create a Multi-Agent Pipeline with kubeswarm

Compose multiple kubeswarm agents into a DAG pipeline using SwarmTeam. Each step targets an agent, passes data via template expressions and tracks execution in a SwarmRun record.

Define agents

apiVersion: kubeswarm.io/v1alpha1
kind: SwarmAgent
metadata:
  name: researcher
spec:
  model: gpt-4o-mini
  prompt:
    inline: "You are a research assistant."
  apiKeyRef:
    name: provider-api-keys
    key: OPENAI_API_KEY
---
apiVersion: kubeswarm.io/v1alpha1
kind: SwarmAgent
metadata:
  name: writer
spec:
  model: claude-sonnet-4-6
  prompt:
    inline: "You are a technical writer."
  apiKeyRef:
    name: provider-api-keys
    key: ANTHROPIC_API_KEY

Define the team

apiVersion: kubeswarm.io/v1alpha1
kind: SwarmTeam
metadata:
  name: blog-pipeline
spec:
  roles:
    - name: researcher
      swarmAgent: researcher
    - name: writer
      swarmAgent: writer

  inputs:
    - name: topic
      type: string
      required: true

  pipeline:
    - role: researcher
      inputs:
        prompt: "Research this topic: {{ .input.topic }}"
    - role: writer
      dependsOn: [researcher]
      inputs:
        prompt: |
          Write a blog post based on this research:
          {{ .steps.researcher.output }}

  output: "{{ .steps.writer.output }}"
  timeoutSeconds: 600

Trigger a run

kubectl apply -f - <<EOF
apiVersion: kubeswarm.io/v1alpha1
kind: SwarmRun
metadata:
  name: blog-run-1
spec:
  teamRef: blog-pipeline
  input:
    topic: "Kubernetes-native agent orchestration"
EOF

kubectl get swrun blog-run-1 -w

Template syntax

Expression	Description
{{ .input.<key> }}	Pipeline input value
{{ .steps.<role>.output }}	Previous step's output
{{ .steps.<role>.artifacts.<name> }}	Step artifact reference