Is create_agent formally producing a ReAct agent?

OSS Product Help LangChain
1

This function seems to replace `create_react_agent` and the mix of system prompt + tools is typical of ReAct agent. Yet one ingredient seems to be missing, from the quickstart doc I am reading: the agent loop. Is it built-in? Is the agent expected to run in loops and can I control this loop? I haven’t found the relevant APIs yet.
This doc seems to imply that there is a loop: Agents - Docs by LangChain
But I don’t see much control over the loop.

2

hi @eric-burel

Yes - create_agent constructs a LangGraph StateGraph that implements a ReAct-style loop (model → tools → model …) and runs until a stop condition is met. The loop is built-in.

How to control the loop:

  • Limit iterations via the LangGraph executor’s recursion_limit config when invoking/streaming the agent:
assert app.invoke(
    {"messages": [HumanMessage(content="what is weather in sf")]},
    {"recursion_limit": 2},
    debug=True,
) == {
    "messages": [
        _AnyIdHumanMessage(content="what is weather in sf"),
        _AnyIdAIMessage(content="Sorry, need more steps to process this request."),
    ]
}
  • Exit early when a tool has return_direct=True or when a structured output tool completes (see code excerpts above).
  • Skip the loop entirely by passing no tools; the graph becomes a single model node (no tool-calling loop).
  • Programmatic control with middleware: set state["jump_to"] to "model" | "tools" | "end" to steer or terminate the loop (see _resolve_jump and the conditional edges around it in the same file).
  • Pause/resume specific nodes with interrupt_before / interrupt_after (useful for human-in-the-loop UIs); these are exposed in create_agent(...).compile(...) arguments.
1 Like