LangGraph is my go-to framework for building graph-based agentic workflows involving multiple agents.

Oftentimes, you need to facilitate a conversation between the user and the LLM and make sure the conversation history is retained for every subsequent calls.

Therefore every time you make a call to the the LLM API, in one way or the other you need to include the messages history in the payload, as a sequence of human and AI messages.

In LangGraph we keep messages in the state, which might look something like:

# MessagesState as defined in langgraph.graph
# class MessagesState(TypedDict):
#  messages: Annotated[list[AnyMessage], add_messages]

# Inherit MessagesState in your app's state class to get access to "messages" and the reducer
class AgentState(MessagesState):
 state_attr: str

Then, just before calling your agent, which is basically an LLM call with a specific prompt (and tools if required), you create the system prompt as below:

from langchain_core.messages import (
 SystemMessage,
 AIMessage,
 get_buffer_string,
)

def agent_node(state: AgentState) -> AgentState:
 AGENT_PROMPT = """You are an expert in this and that.
  
 **Your tasks:**
 - blah
 - blah
 
 These are the messages that have been exchanged so far with the user:
 <Messages>
 {messages}
 </Messages>
 """
 
 SYSTEM_MESSAGE = SystemMessage(
  content=AGENT_PROMPT.format(
    messages=get_buffer_string(state["messages"]),
  )
 )
 
 llm_response = get_agent().invoke([SYSTEM_MESSAGE])
 
 # Update state
 return {
  "messages": [AIMessage(content=llm_response["messages"][-1].content)]
 }

Using get_buffer_string will include the conversation in a way that the LLM usually can interpret better as a series of human and AI messages.

The system prompt will look like:

You are an expert in this and that.
  
**Your tasks:**
- blah
- blah

These are the messages that have been exchanged so far with the user:
<Messages>
Human: ....
AI: ....
Human: ....
AI: ....
Human: ....
AI: ....
Human: ....
AI: ....
</Messages>

As a developer, you get a cleaner approach, that you can log and visually understand as simple sentences in sequence with role, rather than logging the list of message objects.