Skip to content

AI-Human Interaction

This example demonstrates how AI agents can interact with humans:

  • Using agent capabilities for human interaction
  • Setting up a human agent in the pool
  • Allowing AI to request human input when needed

Configuration

Our human_agents.yml sets up an AI assistant and a human agent:

agents:
  assistant:
    model: openai:gpt-4o-mini
    capabilities:
      can_ask_agents: true  # Allow asking other agents (including humans)
    system_prompts:
      - |
        You are a helpful assistant. When you're not sure about something,
        don't hesitate to ask the human agent for guidance.

  human:
    provider: "human"  # Special provider type for human interaction
    description: "A human who can provide answers"

Implementation

Here's how we set up the human-AI interaction:

from llmling_agent.delegation import AgentPool

# A question that requires human knowledge
QUESTION = """
What is the current status of Project DoomsDay?
This is crucial information that only a human would know.
If you don't know, ask the agent named "human".
"""

async def main():
    async with AgentPool[None]("human_agents.yml") as pool:
        assistant = pool.get_agent("assistant")
        # This will cause the assistant to interact with the human
        await assistant.run(QUESTION)
        # Show the complete conversation
        print(await assistant.conversation.format_history())

if __name__ == "__main__":
    import asyncio
    asyncio.run(main())

How It Works

  1. We set up two agents:
  2. An AI assistant with can_ask_agents capability

  3. A human agent using the special "human" provider type

  4. When the AI assistant encounters a question it can't answer:

  5. It recognizes the need for human input
  6. Uses its can_ask_agents capability to interact with the human agent

  7. Incorporates the human's response into its answer

  8. The conversation might look like this: 

    Assistant: I need to check about Project DoomsDay's status. Let me ask the human.
Human: Project DoomsDay is currently in Phase 2, with 60% completion.
Assistant: Based on the human's input, Project DoomsDay is in Phase 2 and is 60% complete.

This demonstrates how to:

  • Enable AI-human collaboration
  • Control when AI can request human input
  • Integrate human knowledge into AI responses