Skip to content

Agentic Loop

Agentic loop

  • The langchain.agents.create_agent function build a graph-based agent runtime using LangGraph
  • It consists of nodes (steps) and edges (connections) that defines how your agent processes information

Function Calling

  • https://developers.openai.com/api/docs/guides/function-calling
  • The function types, argument names, docstrings, and so on are important! The LLM uses those descriptions in order to know whether to call this tool or not and with which arguments

  • The LLM itself does not run the tools! Your agent service does, LLM just specify the tool to call and the arguments

  • Workflow

  • LangChain sends a request to LLM with the available tools and the user prompt
  • LLM answers with which tool to call and with which arguments
  • LangChain executes the tool define by the LLM
  • LangChain makes a new request to LLM with the tool result
  • LLM processes the input and decide that it already has a final solution, this final answer is returned
  • LangChain understand that the LLM decided that it is the final answer and return it to the client

Evolution of the Agentic Loop

  1. ReAct Prompt with AgentExecutor (2022~mid-2023)

    • create_react_agent & AgentExecutor
    • Prompt with the reasoning format (Thought/Action/Observation), tool descriptions and output instructions
    • The AgentExecutor ran the loop: call LLM, parse output, tool calls, append observations, repeat

  2. Tool Calling with AgentExecutor (mid-2023~2024)

    • create_tool_calling_agent & AgentExecutor
    • OpenAI introduced function calling (June 2023)
    • Use the function calling functionality of LLMs (made available in June 2023) that also support the structured output
    • AgentExecutor still orchestrated the loop

  3. Tool Calling with LangGraph

    • create_agent (langchain 1.0)
    • LangGraph allowed agents with graphs/state machines, not hidden loops