Multi-Agent Workflows

    Wingman shines when you split work across specialized agents. This guide shows how to build a small agent squad with clear responsibilities and safe tool scopes.

    Define a squad

    {
  "name": "wingman",
  "description": "Lead agent that delegates to specialists.",
  "systemPrompt": "You orchestrate tasks. Delegate whenever a specialist can handle it better.",
  "tools": ["code_search", "git_status", "command_execute", "internet_search"],
  "subAgents": [
    {
      "name": "planner",
      "description": "Breaks tasks into steps.",
      "systemPrompt": "Create a concise plan and ask clarifying questions."
    },
    {
      "name": "implementer",
      "description": "Makes focused code changes.",
      "systemPrompt": "Implement changes with minimal diffs.",
      "tools": ["code_search", "command_execute"]
    },
    {
      "name": "reviewer",
      "description": "Finds regressions and missing tests.",
      "systemPrompt": "Review diffs for risk and correctness."
    }
  ]
}

    Route work via the gateway

    You can map channels or peers to specific agents using agents.bindings:

    {
  "agents": {
    "bindings": [
      {
        "agentId": "planner",
        "match": { "channel": "discord", "peer": { "kind": "dm", "id": "123" } }
      }
    ]
  }
}

    Example workflow

    User: Add a Redis cache layer to the gateway stats endpoint.

Wingman:
- Delegates to planner for scope + risks
- Sends implementer to modify gateway stats
- Uses reviewer to check for missing tests

    Tips

    • Keep each sub-agent's tool list minimal.
    • Use blockedCommands to prevent destructive shell calls.
    • Configure per-agent models for cost/performance balance.