SwarmKit: governed multi-agent AI, defined as data

Agents you can actually put in front of operations

Building multi-agent systems usually means hundreds of lines of orchestration code per topology, with governance bolted on afterwards. Change the structure and you're refactoring code.

SwarmKit makes the structure configuration and makes governance part of the runtime. That's what lets AI agents work near production systems: every action is permission-checked, every approval that belongs to a human stays with a human, and every step is on the record.

apiVersion: swarmkit/v1
kind: Topology
metadata:
  name: code-review
agents:
  root:
    id: supervisor
    archetype: supervisor-leader
    children:
      - id: code-reviewer
        archetype: code-analyst
        skills: [code-quality-review, security-scan]
      - id: test-analyst
        archetype: test-analyst
        skills: [test-coverage-review, run-tests]

What's in the box

• Topologies as YAML, compiled to a LangGraph state graph at load time
• Governance built in: IAM scopes, a policy engine, and human-only approval scopes agents cannot be granted
• Quality gates: validation, LLM-as-judge and multi-party human approval in one reusable funnel
• Hash-chained audit trail of every step and tool call
• MCP tool integration configured in YAML, with sandboxing
• Coding-agent nodes: run Claude Code or other harnesses as governed nodes in isolated worktrees
• Many model providers, including any OpenAI- or Ollama-compatible endpoint
• Served over CLI, HTTP, MCP and A2A, with a fleet panel UI
• Well tested: thousands of automated tests in CI