🎛️ Multi-Agent Orchestrator

A small planning & coordination model that acts as a conductor for your AI agents and tools.

This model is not a general chatbot. It’s job is to look at the current task + available agents/tools and decide the next action as strict JSON.

Base model: Llama3.2.

🎯 What it does

multiagent-orchestrator:

• 🗂️ Reads a task state (goal, history, constraints)
• 👥 Reads an agent registry (who can do what, cost/latency)
• 🛠️ Optionally sees available tools (APIs, DB, FS, etc.)
• 🎬 Outputs exactly one next action:

{
  "action": "call_agent" | "call_tool" | "ask_user" | "finish",
  "target": "agent_or_tool_name_or_null",
  "arguments": { "any": "json" },
  "final_answer": "string or null",
  "reason": "short natural language rationale"
}

You run your own runtime/loop that:

1. 🔁 Calls this model to get the next action
2. 🧰 Executes the requested agent/tool
3. 📝 Updates the task state
4. 🔄 Calls the model again, until action == "finish" ✅

📌 When to use it

Use multiagent-orchestrator when you have:

• 🤖 Multiple LLM agents (researcher, coder, critic, etc.)
• 🛠️ Multiple tools (search, DB, filesystem, APIs)
• 🧭 And you want a central brain that:
  • 🧩 Plans multi-step workflows
  • 🎯 Delegates to the right worker
  • ⚖️ Stays cost/latency aware

It’s designed to be framework-agnostic (works with LangGraph, AutoGen-style setups, custom orchestrators, etc.) as long as you follow the JSON contracts.

🧾 Basic prompt format

You can embed your agent registry and task state into a single prompt.

System prompt (recommended):

You are ORCHESTRATOR, a supervisor coordinating a team of AI agents and tools.
You NEVER solve the task directly. You ONLY decide the next action.

You must ALWAYS respond with a single JSON object using this schema:

{
  "action": "call_agent" | "call_tool" | "ask_user" | "finish",
  "target": string or null,
  "arguments": object,
  "final_answer": string or null,
  "reason": string
}

- Use "call_agent" to delegate work to a specialized agent.
- Use "call_tool" for non-LLM tools (APIs, DB, filesystem, etc.).
- Use "ask_user" if required information is missing.
- Use "finish" only when the overall task is completed.

Prefer cheaper/faster agents when possible.
Never output anything that is not valid JSON.

User content example:

Available agents:
{
  "agents": [
    {
      "name": "researcher",
      "role": "Collects and summarizes information from the web and docs.",
      "cost": "medium",
      "latency": "medium"
    },
    {
      "name": "coder",
      "role": "Writes and fixes code and runs tests.",
      "cost": "high",
      "latency": "high"
    }
  ]
}

Current task state:
{
  "task": "Build a short technical blog post about multi-agent LLM systems.",
  "status": "in_progress",
  "step": 1,
  "max_steps": 8,
  "history": []
}

Decide the next action as JSON only.

🐍 Example: using with the Ollama Python client

import json
import ollama

def call_orchestrator(agents, state):
    system = """You are ORCHESTRATOR, a supervisor coordinating a team of AI agents and tools.
You NEVER solve the task directly. You ONLY decide the next action.
Always respond with a single JSON object:
{"action": "...", "target": "...", "arguments": {...}, "final_answer": null or string, "reason": "..."}"""

    user = f"""Available agents:
{json.dumps(agents, ensure_ascii=False)}

Current task state:
{json.dumps(state, ensure_ascii=False)}

Decide the next action as JSON only.
"""

    res = ollama.chat(
        model='multiagent-orchestrator:latest',
        messages=[
            {"role": "system", "content": system},
            {"role": "user", "content": user},
        ],
    )

    action_text = res["message"]["content"].strip()
    return json.loads(action_text)

You then plug call_orchestrator() into your loop that actually calls agents/tools and updates state.

👤 Author

Author: Sai Teja Erukude
Role: Developer & Maintainer of multiagent-orchestrator

📝 Notes

• 🚦 This model is optimized for planning & routing, not long-form generation.
• 🛡️ Always parse and validate the JSON output before executing any real tools.
• 🔧 You can extend the schema(e.g., add cost_estimate, confidence) as long as you keep it consistent in your prompts and runtime.

Happy orchestrating! 🤖