On this article, you’ll find out how the ReAct (Reasoning + Performing) sample works and the way to implement it with LangGraph — first with a easy, hardcoded loop after which with an LLM-driven agent.
Subjects we’ll cowl embrace:
- The ReAct cycle (Motive → Act → Observe) and why it’s helpful for brokers.
- Easy methods to mannequin agent workflows as graphs with LangGraph.
- Constructing a hardcoded ReAct loop, then upgrading it to an LLM-powered model.
Let’s discover these methods.
Constructing ReAct Brokers with LangGraph: A Newbie’s Information
Picture by Creator
What’s the ReAct Sample?
ReAct (Reasoning + Performing) is a standard sample for constructing AI brokers that assume by issues and take actions to resolve them. The sample follows a easy cycle:
- Reasoning: The agent thinks about what it must do subsequent.
- Performing: The agent takes an motion (like looking for data).
- Observing: The agent examines the outcomes of its motion.
This cycle repeats till the agent has gathered sufficient data to reply the person’s query.
Why LangGraph?
LangGraph is a framework constructed on prime of LangChain that allows you to outline agent workflows as graphs. A graph (on this context) is an information construction consisting of nodes (steps in your course of) related by edges (the paths between steps). Every node within the graph represents a step in your agent’s course of, and edges outline how data flows between steps. This construction permits for complicated flows like loops and conditional branching. For instance, your agent can cycle between reasoning and motion nodes till it gathers sufficient data. This makes complicated agent habits simple to grasp and preserve.
Tutorial Construction
We’ll construct two variations of a ReAct agent:
- Half 1: A easy hardcoded agent to grasp the mechanics.
- Half 2: An LLM-powered agent that makes dynamic selections.
Half 1: Understanding ReAct with a Easy Instance
First, we’ll create a primary ReAct agent with hardcoded logic. This helps you perceive how the ReAct loop works with out the complexity of LLM integration.
Setting Up the State
Each LangGraph agent wants a state object that flows by the graph nodes. This state serves as shared reminiscence that accumulates data. Nodes learn the present state and add their contributions earlier than passing it alongside.
|
from langgraph.graph import StateGraph, END from typing import TypedDict, Annotated import operator
# Outline the state that flows by our graph class AgentState(TypedDict): messages: Annotated[list, operator.add] next_action: str iterations: int |
Key Parts:
StateGraph: The principle class from LangGraph that defines our agent’s workflow.AgentState: A TypedDict that defines what data our agent tracks.messages: Makes use ofoperator.addto build up all ideas, actions, and observations.next_action: Tells the graph which node to execute subsequent.iterations: Counts what number of reasoning cycles we’ve accomplished.
Making a Mock Device
In an actual ReAct agent, instruments are capabilities that carry out actions on the planet — like looking out the net, querying databases, or calling APIs. For this instance, we’ll use a easy mock search device.
|
# Easy mock search device def search_tool(question: str) -> str: # Simulate a search – in actual utilization, this might name an API responses = { “climate tokyo”: “Tokyo climate: 18°C, partly cloudy”, “inhabitants japan”: “Japan inhabitants: roughly 125 million”, } return responses.get(question.decrease(), f“No outcomes discovered for: {question}”) |
This operate simulates a search engine with hardcoded responses. In manufacturing, this might name an actual search API like Google, Bing, or a customized information base.
The Reasoning Node — The “Mind” of ReAct
That is the place the agent thinks about what to do subsequent. On this easy model, we’re utilizing hardcoded logic, however you’ll see how this turns into dynamic with an LLM in Half 2.
|
# Reasoning node – decides what to do def reasoning_node(state: AgentState): messages = state[“messages”] iterations = state.get(“iterations”, 0)
# Easy logic: first search climate, then inhabitants, then end if iterations == 0: return {“messages”: [“Thought: I need to check Tokyo weather”], “next_action”: “motion”, “iterations”: iterations + 1} elif iterations == 1: return {“messages”: [“Thought: Now I need Japan’s population”], “next_action”: “motion”, “iterations”: iterations + 1} else: return {“messages”: [“Thought: I have enough info to answer”], “next_action”: “finish”, “iterations”: iterations + 1} |
The way it works:
The reasoning node examines the present state and decides:
- Ought to we collect extra data? (return
"motion") - Do now we have sufficient to reply? (return
"finish")
Discover how every return worth updates the state:
- Provides a “Thought” message explaining the choice.
- Units
next_actionto path to the subsequent node. - Increments the iteration counter.
This mimics how a human would method a analysis process: “First I would like climate information, then inhabitants knowledge, then I can reply.”
The Motion Node — Taking Motion
As soon as the reasoning node decides to behave, this node executes the chosen motion and observes the outcomes.
|
# Motion node – executes the device def action_node(state: AgentState): iterations = state[“iterations”]
# Select question primarily based on iteration question = “climate tokyo” if iterations == 1 else “inhabitants japan” consequence = search_tool(question)
return {“messages”: [f“Action: Searched for ‘{query}'”, f“Observation: {result}”], “next_action”: “reasoning”}
# Router – decides subsequent step def route(state: AgentState): return state[“next_action”] |
The ReAct Cycle in Motion:
- Motion: Calls the search_tool with a question.
- Statement: Data what the device returned.
- Routing: Units next_action again to “reasoning” to proceed the loop.
The router operate is a straightforward helper that reads the next_action worth and tells LangGraph the place to go subsequent.
Constructing and Executing the Graph
Now we assemble all of the items right into a LangGraph workflow. That is the place the magic occurs!
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 |
# Construct the graph workflow = StateGraph(AgentState) workflow.add_node(“reasoning”, reasoning_node) workflow.add_node(“motion”, action_node)
# Outline edges workflow.set_entry_point(“reasoning”) workflow.add_conditional_edges(“reasoning”, route, { “motion”: “motion”, “finish”: END }) workflow.add_edge(“motion”, “reasoning”)
# Compile and run app = workflow.compile()
# Execute consequence = app.invoke({“messages”: [“User: Tell me about Tokyo and Japan”], “iterations”: 0, “next_action”: “”})
# Print the dialog movement print(“n=== ReAct Loop Output ===”) for msg in consequence[“messages”]: print(msg) |
Understanding the Graph Construction:
- Add Nodes: We register our reasoning and motion capabilities as nodes.
- Set Entry Level: The graph at all times begins on the reasoning node.
- Add Conditional Edges: Primarily based on the reasoning node’s determination:
- If
next_action == "motion"→ go to the motion node. - If
next_action == "finish"→ cease execution.
- If
- Add Mounted Edge: After motion completes, at all times return to reasoning.
The app.invoke() name kicks off this whole course of.
Output:
|
=== ReAct Loop Output === Consumer: Inform me about Tokyo and Japan
Thought: I want to verify Tokyo climate Motion: search(‘climate tokyo’) Statement: Tokyo climate: 18°C, partly cloudy
Thought: Now I want Japan‘s inhabitants Motion: search(‘inhabitants japan‘) Statement: Japan inhabitants: roughly 125 million
Thought: I have sufficient information to reply |
Now let’s see how LLM-powered reasoning makes this sample actually dynamic.
Half 2: LLM-Powered ReAct Agent
Now that you just perceive the mechanics, let’s construct a actual ReAct agent that makes use of an LLM to make clever selections.
Why Use an LLM?
The hardcoded model works, however it’s rigid — it may possibly solely deal with the precise situation we programmed. An LLM-powered agent can:
- Perceive various kinds of questions.
- Determine dynamically what data to assemble.
- Adapt its reasoning primarily based on what it learns.
Key Distinction
As an alternative of hardcoded if/else logic, we’ll immediate the LLM to determine what to do subsequent. The LLM turns into the “reasoning engine” of our agent.
Setting Up the LLM Setting
We’ll use OpenAI’s GPT-4o as our reasoning engine, however you can use any LLM (Anthropic, open-source fashions, and so on.).
|
from langgraph.graph import StateGraph, END from typing import TypedDict, Annotated import operator import os from openai import OpenAI
consumer = OpenAI(api_key=os.environ.get(“OPENAI_API_KEY”))
class AgentStateLLM(TypedDict): messages: Annotated[list, operator.add] next_action: str iteration_count: int |
New State Definition:
AgentStateLLM is much like AgentState, however we’ve renamed it to differentiate between the 2 examples. The construction is equivalent — we nonetheless monitor messages, actions, and iterations.
The LLM Device — Gathering Data
As an alternative of a mock search, we’ll let the LLM reply queries utilizing its personal information. This demonstrates how one can flip an LLM right into a device!
|
def llm_tool(question: str) -> str: “”“Let the LLM reply the question straight utilizing its information”“” response = consumer.chat.completions.create( mannequin=“gpt-4o”, max_tokens=150, messages=[{“role”: “user”, “content”: f“Answer this query briefly: {query}”}] ) return response.decisions[0].message.content material.strip() |
This operate makes a easy API name to GPT-4 with the question. The LLM responds with factual data, which our agent will use in its reasoning.
Observe: In manufacturing, you would possibly mix this with internet search, databases, or different instruments for extra correct, up-to-date data.
LLM-Powered Reasoning — The Core Innovation
That is the place ReAct actually shines. As an alternative of hardcoded logic, we immediate the LLM to determine what data to assemble subsequent.
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 |
def reasoning_node_llm(state: AgentStateLLM): iteration_count = state.get(“iteration_count”, 0) if iteration_count >= 3: return {“messages”: [“Thought: I have gathered enough information”], “next_action”: “finish”, “iteration_count”: iteration_count}
historical past = “n”.be a part of(state[“messages”]) immediate = f“”“You might be an AI agent answering: “Inform me about Tokyo and Japan“
Dialog up to now: {historical past}
Queries accomplished: {iteration_count}/3
You MUST make precisely 3 queries to assemble data. Reply ONLY with: QUERY:
Do NOT be conversational. Do NOT thank the person. ONLY output: QUERY:
determination = consumer.chat.completions.create( mannequin=“gpt-4o”, max_tokens=100, messages=[{“role”: “user”, “content”: prompt}] ).decisions[0].message.content material.strip()
if determination.startswith(“QUERY:”): return {“messages”: [f“Thought: {decision}”], “next_action”: “motion”, “iteration_count”: iteration_count} return {“messages”: [f“Thought: {decision}”], “next_action”: “finish”, “iteration_count”: iteration_count} |
How This Works:
- Context Constructing: We embrace the dialog historical past so the LLM is aware of what’s already been gathered.
- Structured Prompting: We give clear directions to output in a selected format (
QUERY:). - Iteration Management: We implement a most of three queries to stop infinite loops.
- Resolution Parsing: We verify if the LLM needs to take motion or end.
The Immediate Technique:
The immediate tells the LLM:
- What query it’s attempting to reply
- What data has been gathered up to now
- What number of queries it’s allowed to make
- Precisely the way to format its response
- To not be conversational
LLMs are educated to be useful and chatty. For agent workflows, we want concise, structured outputs. This directive retains responses targeted on the duty.
Executing the Motion
The motion node works equally to the hardcoded model, however now it processes the LLM’s dynamically generated question.
|
def action_node_llm(state: AgentStateLLM): last_thought = state[“messages”][–1] question = last_thought.substitute(“Thought: QUERY:”, “”).strip() consequence = llm_tool(question) return {“messages”: [f“Action: query(‘{query}’)”, f“Observation: {result}”], “next_action”: “reasoning”, “iteration_count”: state.get(“iteration_count”, 0) + 1} |
The Course of:
- Extract the question from the LLM’s reasoning (eradicating the “Thought: QUERY:” prefix).
- Execute the question utilizing our llm_tool.
- Report each the motion and remark.
- Route again to reasoning for the subsequent determination.
Discover how that is extra versatile than the hardcoded model — the agent can ask for any data it thinks is related!
Constructing the LLM-Powered Graph
The graph construction is equivalent to Half 1, however now the reasoning node makes use of LLM intelligence as an alternative of hardcoded guidelines.
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 |
workflow_llm = StateGraph(AgentStateLLM) workflow_llm.add_node(“reasoning”, reasoning_node_llm) workflow_llm.add_node(“motion”, action_node_llm) workflow_llm.set_entry_point(“reasoning”) workflow_llm.add_conditional_edges(“reasoning”, lambda s: s[“next_action”], {“motion”: “motion”, “finish”: END}) workflow_llm.add_edge(“motion”, “reasoning”)
app_llm = workflow_llm.compile() result_llm = app_llm.invoke({ “messages”: [“User: Tell me about Tokyo and Japan”], “next_action”: “”, “iteration_count”: 0 })
print(“n=== LLM-Powered ReAct (No Mock Information) ===”) for msg in result_llm[“messages”]: print(msg) |
What’s Completely different:
- Similar graph topology (reasoning ↔ motion with conditional routing).
- Similar state administration method.
- Solely the reasoning logic modified – from if/else to LLM prompting.
This demonstrates the facility of LangGraph: you may swap parts whereas retaining the workflow construction intact!
The Output:
You’ll see the agent autonomously determine what data to assemble. Every iteration exhibits:
- Thought: What the LLM determined to ask about.
- Motion: The question being executed.
- Statement: The knowledge gathered.
Watch how the LLM strategically gathers data to construct a whole reply!
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 |
=== LLM–Powered ReAct (No Mock Information) === Consumer: Inform me about Tokyo and Japan
Thought: QUERY: What is the historical past and significance of Tokyo in Japan?
Motion: question(‘What’s the historical past and significance of Tokyo in Japan?’)
Statement: Tokyo, initially identified as Edo, has a wealthy historical past and important function in Japan. It started as a small fishing village till Tokugawa Ieyasu established it as the heart of his shogunate in 1603, marking the begin of the Edo interval. Throughout this time, Edo flourished as a political and cultural hub, changing into one of the world‘s largest cities by the 18th century.
In 1868, after the Meiji Restoration, the emperor moved from Kyoto to Edo, renaming it Tokyo, that means “Japanese Capital”. This transformation marked the start of Tokyo’s modernization and fast growth. Over the twentieth century, Tokyo confronted challenges, together with the Nice Kanto Earthquake in 1923 and heavy bombings
Thought: QUERY: What are the main cultural and financial contributions of Tokyo to Japan?
Motion: question(‘What are the foremost cultural and financial contributions of Tokyo to Japan?’)
Statement: Tokyo, as the capital of Japan, is a main cultural and financial powerhouse. Culturally, Tokyo is a hub for conventional and up to date arts, together with theater, music, and visible arts. The metropolis is residence to quite a few museums, galleries, and cultural websites such as the Tokyo Nationwide Museum, Senso–ji Temple, and the Meiji Shrine. It additionally hosts worldwide occasions like the Tokyo Worldwide Movie Pageant and numerous style weeks, contributing to its status as a world style and cultural heart.
Economically, Tokyo is one of the world‘s main monetary facilities. It hosts the Tokyo Inventory Alternate, one of many largest inventory exchanges globally, and is the headquarters for quite a few multinational companies. Town’s superior infrastructure and innovation in know-how and trade make it a focal
Thought: QUERY: What are the key historic and cultural points of Japan as a entire?
Motion: question(‘What are the important thing historic and cultural points of Japan as an entire?’)
Statement: Japan boasts a wealthy tapestry of historic and cultural points, formed by centuries of growth. Traditionally, Japan‘s tradition was influenced by its isolation as an island nation, main to a distinctive mix of indigenous practices and international influences. Key historic intervals embrace the Jomon and Yayoi eras, characterised by early settlement and tradition, and the subsequent intervals of imperial rule and samurai governance, such as the Heian, Kamakura, and Edo intervals. These intervals fostered developments like the tea ceremony, calligraphy, and kabuki theater.
Culturally, Japan is identified for its Shinto and Buddhist traditions, which coexist seamlessly. Its aesthetic ideas emphasize simplicity and nature, mirrored in conventional structure, gardens, and arts such as ukiyo–e prints and later
Thought: I have gathered sufficient data |
Wrapping Up
You’ve now constructed two ReAct brokers with LangGraph — one with hardcoded logic to be taught the mechanics, and one powered by an LLM that makes dynamic selections.
The important thing perception? LangGraph allows you to separate your workflow construction from the intelligence that drives it. The graph topology stayed the identical between Half 1 and Half 2, however swapping hardcoded logic for LLM reasoning reworked a inflexible script into an adaptive agent.
From right here, you may lengthen these ideas by including actual instruments (internet search, calculators, databases), implementing device choice logic, and even constructing multi-agent methods the place a number of ReAct brokers collaborate.
