Why langgraph ReAct agent does not have a 'Thought' step?

[fengsp]

Example Code
from langgraph.prebuilt import create_react_agent

Description
When this agent runs, it never show any 'Thought' step

[hwchase17]

it is using function calling directly, instead of the thought/action/observation prompting style

[hinthornw]

To add to this, some models like Claude 3 perform function calling with a "thought" step built-in, while others (like OpenAI) do not. It's a bit provider-dependent right now

[emanueleielo]

it is using function calling directly, instead of the thought/action/observation prompting style

Hi ,I’ve actually worked on this and implemented a think tool in a LangGraph ReAct agent, inspired by Anthropic’s Claude 3 “think” tool described in their engineering blog.

It adds a custom think() function that acts like a structured scratchpad for intermediate reasoning before tool calls. I’ve shared a working example here:

Repo: github.com/emanueleielo/langgraph-think-tool

Would love feedback or suggestions for improving it!

[itxabdullahbinasad]

@hwchase17 Is there a way we can add it in the prompt with adding the variables of tool names and tool in the prompt and just telling it in prompt but then should i give this prompt to the function tool calling agent or langgraphs prebuilt create react agent in the message modifier

[hwchase17]

yes, you could definitely add this in - either as a field in the tool calls or by telling it think before answering (models like GPT-4o can generate text before a tool call)

[kashyap-aditya]

Can we get some documentation on how to get the Thought/Observation/Action pair to work along with tool calling?

Are you suggesting that we add Thought/Observation as input parameters to each tool?

[emanueleielo]

Check it https://github.com/emanueleielo/langgraph-think-tool

[gianmario-msl]

@emanueleielo thanks for sharing this think tool implementation. I am wondering why there is a second llm call inside the tool implementation. The ReAct agent already produced the thought as tool argument we could just return that as AI message or to save tokens we could just return a message like "Thought successfully created" since the original thought message is already in the chat history as part of the Tool message.

Could you elaborate more on the need of invoking an llm with the think prompt? Thanks

[emanueleielo]

@emanueleielo thanks for sharing this think tool implementation. I am wondering why there is a second llm call inside the tool implementation. The ReAct agent already produced the thought as tool argument we could just return that as AI message or to save tokens we could just return a message like "Thought successfully created" since the original thought message is already in the chat history as part of the Tool message.

Could you elaborate more on the need of invoking an llm with the think prompt? Thanks

Actually you are right, I followed your suggestion and removed the LLM call, this is more efficient, but then I had a problem because the model was answering always with the "thought" and not using it as just a scratchpad.

Editing the system prompt fixed it, check it and let me know what do you think about it.

Thank u

[luizzappa]

With OpenAI, tools are called without "thought". The workaround I did was to add a "reason" field to each available tool in which LLM must fill in the reason for calling that function, thus forcing the intermediate thoughts to be recorded. But it is worth testing whether this improves performance in your use case or not.

[gianmario-msl]

I guess that would help with explaining "why calling this tool" although by packing it with the tool arguments means the model already decided to use that tool.
While having an explicit think step before invoking tool can help with reasoning about current observations and decide which tool to call next.
Moreover, when multiple tools are called in parallel you may have redundant thought messages.

[kashyap-aditya]

I am not sure it is accurate to call it a react agent when there is no Thought/Observation/Action pair being output.

What makes it a react agent and how is it different from just a tool calling agent?

[hinthornw]

It's just an homage to original paper that made the looping language model call pattern popular.

If you want to debate semantics then sure, you're correct. It's not literally text-davinci-002 being provided few-shot strings of thought: action: observation.

It's also not literally "calling" things, especially not "tools". And you can debate whether a random seed plus text plus matrix multiplications have any real "agency".

[gianmario-msl]

I think an efficient and elegant implementation could be to have a think tool defined as simple as a Pydantic model class with the thought field (or depending on the domain more specific fields) and to instruct the agent in the system prompt to always call the think tool along with other tools at each iteration. By calling it together with the other tools (as opposed to calling it in isolation in between tool calling steps) reduces the number of model invocations thus saving tokens and reducing latency.

Thus, the current ReAct API does not need to be modified as long as you add this think tool and provide prompt instructions.

[maoismiao]

I've already tried it, and the results are great, but there are requirements for the models. Both Gemini 2.5 Flash and GPT-4.1 work fine.
prompt：https://gist.github.com/maoismiao/f691229d921a69af713133286e264079

[zhangyanbo2007]

your code when call tools it's ok ,but when reply result,there is no think tool to start.can you check it?

[hinthornw]

don't even need pydantic.

def think(thought: str):
    """Plan out your next steps. Takes no action."""
    return ""

[sesiii]

Why "create_react_agent" creates a ReAct agent in langchain but is from langgraph library?
Is there something i am missing?

[TzviPM]

langgraph is apart of the langchain exosystem. Models are coming from langchain, but being orchestrated in langgraph