AgentLauncher

Contact Me

• Blog: https://cugtyt.github.io/blog/llm-application/index
• Email: cugtyt@qq.com
• GitHub: Cugtyt@GitHub

---

AgentLauncher GitHub: https://github.com/Cugtyt/agentlauncher

AgentLauncher is an event-driven, multi-agent framework for solving complex tasks by dynamically generating sub-agents. The main agent coordinates strategy, while sub-agents handle specialized tasks. Agent lifecycles are managed automatically, similar to jobs in Kubernetes—sub-agents are lightweight and ephemeral.

How It Works

1. Task Initialization: Launching AgentLauncher with a task triggers a TaskCreateEvent.
2. Agent Creation: AgentRuntime responds and triggers AgentCreateEvent.
3. Main Agent Management: AgentManager creates the main agent (if needed) and triggers AgentStartEvent.
4. Agent Startup: The agent logs the initial message (MessageAddEvent) and requests an LLM response (LLMRequestEvent).
5. LLM Interaction: LLMRuntime calls the LLM API, triggering LLMResponseEvent.
6. Message Handling: The agent logs the LLM response and may request tool execution (ToolsExecRequestEvent).
7. Tool Execution: ToolRuntime executes tools, triggering ToolExecStartEvent, ToolExecFinishEvent, and ToolsExecResultsEvent.
8. Result Processing: The agent processes tool results and may request further LLM responses, repeating the cycle.
9. Task Completion: The agent triggers AgentFinishEvent. AgentRuntime marks completion, triggers TaskFinishEvent, and cleans up sub-agents.
10. Final Output: AgentLauncher returns the result. New tasks restart the flow.

Features

• Multi-Agent System: Main agent delegates to sub-agents; standardized lifecycle and interactions.
• Event-Driven Architecture: All agent behavior is event-based.
• Dynamic Agent Management: Agents spawn sub-agents for specialized tasks; lifecycles are automatic.
• Modular & Extensible: Easily add new tools and LLM handlers by subscribing to events.
• Fully Asynchronous: Built on asyncio for efficient, non-blocking event handling.

Example Usage

See main.py for a usage example. Register tools and LLM handlers, then run tasks interactively:

from agentlauncher import AgentLauncher

launcher = AgentLauncher()

await launcher.register_tool(
    name="calculate",
    function=lambda a, b, c: a * b + c,
    description="Calculate a * b + c.",
    parameters={
        "type": "object",
        "properties": {
            "a": {"type": "integer", "description": "First integer."},
            "b": {"type": "integer", "description": "Second integer."},
            "c": {"type": "integer", "description": "Third integer."},
        },
        "required": ["a", "b", "c"],
    },
)

await launcher.register_main_agent_llm_handler(
    name="gpt-4",
    function=your_llm_function,
)

while True:
    task = input("Enter your task (or 'exit' to quit): ")
    if task.lower() == "exit":
        break
    result = await launcher.run(task)
    print(f"Final Result: {result}")