Write a function to implement a custom LangChain agent.

What you need to know

Three levels of "custom"

Pick the lowest level that works. Middleware keeps you on the maintained runtime; a custom graph means you own more code.

The loop by hand

1from langchain.messages import HumanMessage, ToolMessage23def run_agent(model, tools, question: str, max_steps: int = 5) -> str:4    bound = model.bind_tools(tools)5    by_name = {t.name: t for t in tools}6    messages = [HumanMessage(question)]7    for _ in range(max_steps):8        ai = bound.invoke(messages)9        messages.append(ai)10        if not ai.tool_calls:                 # no tool call = final answer11            return ai.content12        for call in ai.tool_calls:13            result = by_name[call["name"]].invoke(call["args"])14            messages.append(ToolMessage(str(result), tool_call_id=call["id"]))15    return "Stopped: step limit reached."

Every ToolMessage must carry the tool_call_id of the call it answers; providers reject a history where a tool call has no matching result. This loop has no persistence, streaming, error handling or parallel execution, which is why it is for learning, not production.

The same agent as a LangGraph graph

1from langgraph.graph import StateGraph, MessagesState, START, END2from langgraph.prebuilt import ToolNode, tools_condition34def build_custom_agent(model, tools, checkpointer=None):5    bound = model.bind_tools(tools)67    def call_model(state: MessagesState):8        return {"messages": [bound.invoke(state["messages"])]}910    g = StateGraph(MessagesState)11    g.add_node("model", call_model)12    g.add_node("tools", ToolNode(tools))13    g.add_edge(START, "model")14    g.add_conditional_edges("model", tools_condition)   # "tools" or END15    g.add_edge("tools", "model")16    return g.compile(checkpointer=checkpointer)

MessagesState is a state with one messages list whose reducer appends new messages. ToolNode runs all tool calls from the last message, in parallel, and returns ToolMessages. tools_condition routes to "tools" if there are tool calls, otherwise to END. Now you can add your own nodes: for example a check_answer node between the model and END that verifies every price in the answer came from a tool result.

Customising with middleware instead

1from langchain.agents.middleware import dynamic_prompt, ModelRequest23@dynamic_prompt4def store_prompt(request: ModelRequest) -> str:5    return f"You are the store assistant. Store hours today: {hours()}."

Pass it as create_agent(..., middleware=[store_prompt]). The system prompt is rebuilt on every model call without adding anything to the saved message history. This is how most teams "customise" an agent in LangChain 1.x.

A real-life example

An electronics store's product Q&A assistant sometimes quoted prices from memory instead of the get_price tool, and a few answers showed last month's prices.

A system-prompt rule reduced this but did not stop it. The team rebuilt the agent as a StateGraph with one extra node, verify_prices, that runs after the model's final answer. It extracts every rupee amount with a regex and checks it appears in a get_price tool result in the state. If not, it adds a message — "Price Rs 54,999 was not from a tool; call get_price" — and routes back to the model. Wrong-price answers in the weekly audit went from 11 in 2,000 to zero, at the cost of one extra model call on about 3% of conversations.

Follow-up questions to expect

• "When would you not use create_agent?" — When the flow is not a simple loop: a fixed plan-then-execute, a mandatory validation step, or several agents with shared state.
• "What does tools_condition do?" — It checks the last AI message for tool calls and returns "tools" or END.
• "What is a reducer?" — The function that merges a node's update into state; for messages it appends (and replaces by ID) instead of overwriting.