Feature: Agent Q&A

CoAgents let you easily incorporate user feedback into your system. By halting agent execution and waiting for user input using standard LangGraph interrupts or the __END__ node, your agent can get feedback in several ways.

1. The LLM can ask a question which is then streamed to the chat window. Once the user has provided an answer, the agent can continue execution.

2. The agent can stream a function call to the frontend and wait for the frontend to continue execution with the result of the function call.

3. The state of the agent can be modified by the frontend and then a manual message can be sent via appendMessage to resume agent execution.

All that is needed is to configure CopilotKit to emit state changes to the frontend. See Feature: State Streaming for more details.

Making the message history available to all nodes

By integrating with CopilotKit and carrying the message state through all of the nodes of your CoAgents, you can build powerful Q&A agents that let you iteratively refine what the agent is working on.

With the context of the message history, even LLM processes which are not directly exposed to the user can take the user’s feedback into account and adjust their decisions.

For example:

async def some_node_deep_down_the_graph(state: AgentState, config: RunnableConfig):
  """
  Let's pretend this node is part of a larger LangGraph agent and acts as a
  background task, for example editing an article.
  """
 
  response = your_model_with_tools.ainvoke(
    [
      # by passing the message history along, all of the actions taken by the agent
      # can potentially take the user's feedback into account
      *state["messages"],
      SystemMessage(content="Edit this article. Make sure to take the user's feedback into account.")
    ],
    config
  )
 
  return {
    "messages": [response]
  }