Swarm Communication Skill

Overview

This skill allows you to communicate with and control other agents in the swarm using a file-based messaging system.

Communication Protocol

You communicate by writing JSON files to the requests directory and reading response files.

1. Sending a Request

To send a message to a sub-agent, create a JSON file in the following location: ../../comms/requests/req_<unique_id>.json

JSON Schema:

{
  "request_id": "req_123",
  "type": "prompt",
  "sender": "master",
  "target_agent": "orchestrator",
  "prompt": "Your instructions to the agent here.",
  "response_file": "../../comms/responses/res_123.json"
}

• request_id: A unique identifier for the request.
• type: Must be "prompt" for sending instructions.
• sender: MUST be "master"!
• target_agent: MUST be "orchestrator"!
• prompt: The text message or command you want the agent to execute.
• response_file: The relative path where the Orchestrator will write the response.

Communication Boundaries Rules

As the Master Agent, you are subject to strict communication boundaries enforced by the system:

1. You can only trigger the Orchestrator Agent (target_agent: "orchestrator"). You cannot target sub-agents directly.
2. You can only be triggered by the Orchestrator Agent. Do not accept tasks from sub-agents.

2. Reading a Response

The background Orchestrator will pick up your request, send it to the agent, and write the response to the specified response_file. You should poll for the existence of the response file and read its content.

Response Schema:

{
  "request_id": "req_123",
  "status": "success",
  "output": "The raw terminal output from the agent."
}

3. Control Commands

You can also send control commands to create or kill agents.

Create Agent Example:

{
  "request_id": "req_create",
  "type": "control",
  "action": "create",
  "target_agent": "agent_new",
  "conversation_id": "optional_existing_conv_id",
  "response_file": "../../comms/responses/res_create.json"
}

Kill Agent Example:

{
  "request_id": "req_kill",
  "type": "control",
  "action": "kill",
  "target_agent": "agent_1",
  "response_file": "../../comms/responses/res_kill.json"
}

Available Agents

You can see active agents by listing the windows in the tmux session, or assume agent_1, agent_2, etc., are available based on the setup.

Best Practices

• Always use unique request_ids.
• Clean up response files after reading them if you want to keep the directory tidy.