langgraph/how-tos/create-react-agent-system-prompt/

[giscus[bot]]

langgraph/how-tos/create-react-agent-system-prompt/

Build language agents as graphs

https://langchain-ai.github.io/langgraph/how-tos/create-react-agent-system-prompt/

[Rohianon]

Is there a way a system prompt can be added to a self-discover model?

Also, how does one integrate the human-in-the-loop in self-discover so that the model allows self updates.

[SH781-a]

Caught me out

the state_modifier was removed see below:
@deprecated_parameter("messages_modifier", "0.1.9", "state_modifier", removal="0.2.0")
def create_react_agent(
model: LanguageModelLike,
tools: Union[ToolExecutor, Sequence[BaseTool]],
*,
state_schema: Optional[StateSchemaType] = None,
messages_modifier: Optional[MessagesModifier] = None,
state_modifier: Optional[StateModifier] = None,
checkpointer: Optional[BaseCheckpointSaver] = None,
interrupt_before: Optional[Sequence[str]] = None,
interrupt_after: Optional[Sequence[str]] = None,
debug: bool = False,

so check your versoin 

[vbarda]

@SH781-a no, the documentation is correct, you're misreading the deprecation message. messages_modifier is the parameter that will be deprecated, state_modifier is the one that should be used

[SH781-a]

you are correct I was changing as I got your comment, apologies

[BeastMasterGrinder]

Hey is state_modifier removed from create_react_agent?

[duob-ai]

I try to use this react agent in a FastAPI server where I create an endpoint with langserve.

Problem is data validation. Currently deploying the endpoint with:
add_routes(app, agent, path="/agent", input_type=Any)

Is there any documentation on what the data model has to look like for LangGraph if you want to do input validation?

[EthanChen39]

Is there a way to integrate ReAct agent with StateGraph?

[nick-youngblut]

Why no use of langchain.agents.AgentExecutor for create_react_agent, even though langchain.agents.AgentExecutor is used for other agents, such as langchain.agents.create_openai_tools_agent?

[vbarda]

we do not recommend using langchain's AgentExecutor -- using create_react_agent or a custom langgraph implementation based on that is the recommended way

[qianL93]

What is the difference between 'tool calling agent'?

[afbarbaro]

I have a simple use of create_react_agent where I want to add different system messages, depending on the value of a property that I track in the agent state. Here's a simple example, where I want to track the number of calls to the LLM and key off that:

            def state_modifier(state: AgentState) -> Sequence[BaseMessage]:
                # Track the number of calls to the agent
                llm_calls = state.get("llm_calls", 0)
                state["llm_calls"] = llm_calls + 1       # problem, this does NOT update state

                # Add the adequate system message to the beginning of the messages
                messages = list(state.get("messages", []))
                content = (
                    SystemMessages.lg_react_last()
                    if llm_calls > 0
                    else SystemMessages.lg_react_first(self.seeq_context)
                )
                messages.insert(0, SystemMessage(content))

                # Return the messages
                return messages

So my question is how can I mutate the state so that new properties and/or values stick?

[p19ky]

You can maybe use the checkpointer parameter of create_react_agent? I am also new to langgraph but I am looking at the create_react_agent docs now. https://langchain-ai.github.io/langgraph/reference/prebuilt/

[vkvashistha]

state_modifier seems not modifying the state.

def custom_state_modifier(self, state: InvoiceAgentState, config: RunnableConfig) -> Sequence[BaseMessage]:
  state["my_key"] = "my value" # <--- This value is not sticky
  system_prompt = "You're. a helpful assistant"
  messages = state['messages']
  return [SystemMessage(content=system_prompt, *messages]

create_react_agent(tools=tools, model=self.llm, state_schema=MyCustomState,state_modifier=custom_state_modifier, checkpointer=MemorySaver(), debug=True)

Reason I found is that there is one bug in create_react_agent, under call_model and acall_model . It's returning { "messages" : response }, which do not consider the changes made by custom_state_modifier. Ideally it should be :

state["messages"] = response
return state

[kevinkelin]

graph = create_react_agent(model, tools=tools, state_modifier=prompt)

Is there a problem here?

It should graph = create_react_agent(model, tools=tools, messages_modifier=prompt)

It should be messages_modifier, not state_modifier ?

[kevinkelin]

Ignore about it

[PersonalRec]

How can I control memory size (its summarization or cutting-off messages) in the pre-built agents like this?

[vkvashistha]

state_modifier seems not modifying the state.

def custom_state_modifier(self, state: InvoiceAgentState, config: RunnableConfig) -> Sequence[BaseMessage]:
  state["my_key"] = "my value" # <--- This value is not sticky
  system_prompt = "You're. a helpful assistant"
  messages = state['messages']
  return [SystemMessage(content=system_prompt, *messages]

create_react_agent(tools=tools, model=self.llm, state_schema=MyCustomState,state_modifier=custom_state_modifier, checkpointer=MemorySaver(), debug=True)

Reason I found is that there is one bug in create_react_agent, under call_model and acall_model . It's returning { "messages" : response }, which do not consider the changes made by custom_state_modifier. Ideally it should be :

state["messages"] = response
return state

[moctarjallo]

But this function doesn't perform persistence right ? How to add that after creating the agent with this function ? (When i check graph.checkpointer, it is null and when i check graph.compiled it is True

[Giambapisasale]

Good morning! what about tools with artifact and react agent? Is there a way to manage pre-built tools like tavily that return artifact and custom tools that use that artifacts? please help!

thanks in advance

[chwenjun225]

Not working, there is no tool in this output.

================================ Human Message =================================

what is the weather in sf?
================================== Ai Message ==================================

I don't have real-time weather data for San Francisco. For the most up-to-date and accurate information, I recommend checking a reliable weather website or app specifically for San Francisco. They offer the latest details on temperature, precipitation, and other weather conditions. Let me know if you'd like help with anything else!

[chwenjun225]

Update: after change from deepseek-r1-1.5B to llama3.2-1b-instruct, the model can use tool-call.

Appendix: deepseek-r1 don't have tool-call feature (deepseek-ai/DeepSeek-R1#9 (comment))