1 00:00:03,000 --> 00:00:09,439 Welcome to this video on ReAct: Building Agents that Reason Before Acting. 2 00:00:09,439 --> 00:00:14,520 In this video, you'll explore how React enables step-by-step agent reasoning. You'll 3 00:00:14,520 --> 00:00:18,799 identify key components like thought, action, and observation. 4 00:00:18,799 --> 00:00:22,840 You'll construct a working React agent in LangGraph. 5 00:00:22,840 --> 00:00:27,399 And finally, you'll analyze how the agent reasons through tool calls to generate a final 6 00:00:27,399 --> 00:00:30,280 response. 7 00:00:30,280 --> 00:00:35,840 React agents are designed for complex tasks that require step-by-step reasoning. For example, 8 00:00:35,840 --> 00:00:38,360 you might begin with a system prompt like, 9 00:00:38,360 --> 00:00:43,360 You are a helpful AI assistant that thinks step-by-step and uses tools when needed. 10 00:00:43,360 --> 00:00:47,720 You can also prompt the LLM to respond in the following ways. 11 00:00:47,720 --> 00:00:51,919 Identify the information needed. Use tools when appropriate. 12 00:00:51,919 --> 00:00:57,919 Get clear and useful answers. And finally, walk through its reasoning process. 13 00:00:57,919 --> 00:01:02,880 The system message should instruct the LLM to format its output using a structured pattern. 14 00:01:02,880 --> 00:01:08,400 First, it starts with Thought, where the LLM reasons about what to do next. 15 00:01:08,400 --> 00:01:12,519 Then comes Action, the name of the tool to use. 16 00:01:12,519 --> 00:01:17,519 Next is Action Input, the input provided to the selected tool. 17 00:01:17,519 --> 00:01:21,839 After that is Observation, the result returned by the tool. 18 00:01:21,839 --> 00:01:28,239 And then Final Answer, the complete response to the user based on what the tool returned. 19 00:01:28,239 --> 00:01:32,419 Let's start with a simple example. You're a frequent traveler who owns multiple apartments 20 00:01:32,419 --> 00:01:37,120 in different cities and access to an LLM that can use two tools. 21 00:01:37,120 --> 00:01:39,800 A search tool that provides the current weather. 22 00:01:39,800 --> 00:01:45,059 And a clothing recommendation tool that suggests what to wear based on what's in your closet. 23 00:01:45,059 --> 00:01:48,339 In the block diagram, the LLM is shown in blue. 24 00:01:48,339 --> 00:01:50,940 The green box represents the environment. 25 00:01:50,940 --> 00:01:54,660 The LLM must interact with the environment through these tools. 26 00:01:54,660 --> 00:02:00,459 It reasons step-by-step, uses tool feedback, and iterates toward a final answer. 27 00:02:00,459 --> 00:02:02,980 In this case, what to wear. 28 00:02:02,980 --> 00:02:06,779 Consider the question, what's the weather in Tokyo and what should I wear? 29 00:02:06,779 --> 00:02:11,179 The LLM must use tools to obtain data from the environment, in this case, the current 30 00:02:11,179 --> 00:02:16,220 weather, and check what clothes are available using the clothing recommendation tool. 31 00:02:16,220 --> 00:02:20,539 The LLM outputs the thought, I need to look up the weather in Tokyo first. 32 00:02:20,539 --> 00:02:24,940 The LLM determines the action, in this case, to use the search tool. 33 00:02:24,940 --> 00:02:30,300 At the same time, the LLM generates the action input, Tokyo weather today. 34 00:02:30,300 --> 00:02:34,460 This serves as input to the tool, which in turn interacts with the environment to retrieve 35 00:02:34,460 --> 00:02:36,440 the relevant information. 36 00:02:36,440 --> 00:02:39,860 The tool returns an observation after accessing the environment. 37 00:02:39,860 --> 00:02:44,139 Tokyo is 22°C with sunny skies. 38 00:02:44,139 --> 00:02:47,899 The observation and chat history are passed back into the LLM. 39 00:02:47,899 --> 00:02:51,979 The chat history is always included, even if not shown explicitly. 40 00:02:51,979 --> 00:02:55,220 The LLM takes the new input and processes a thought. 41 00:02:55,220 --> 00:02:59,059 Now I should recommend clothing for 22°C sunny weather. 42 00:02:59,059 --> 00:03:00,820 Next, the action. 43 00:03:00,820 --> 00:03:04,059 It selects the tool, recommend_clothing. 44 00:03:04,059 --> 00:03:08,779 The action input is produced, 22°C sunny weather. 45 00:03:08,779 --> 00:03:14,100 After the observation, light clothing recommended, t-shirt, shorts, sunglasses. 46 00:03:14,100 --> 00:03:17,419 Finally, the LLM reaches its last thought. 47 00:03:17,419 --> 00:03:20,059 It now has all the information it needs. 48 00:03:20,059 --> 00:03:21,940 It provides the final answer. 49 00:03:21,940 --> 00:03:25,339 Tokyo is 22°C and sunny today. 50 00:03:25,339 --> 00:03:30,059 I recommend wearing light clothing like a t-shirt, shorts, and sunglasses. 51 00:03:30,059 --> 00:03:33,779 Since there are no further tool calls, the process ends here. 52 00:03:33,779 --> 00:03:37,259 Now let's see how to implement this in LangGraph. 53 00:03:37,259 --> 00:03:41,979 The code example you are going to look at emphasizes simplicity by minimizing the schema, 54 00:03:41,979 --> 00:03:44,619 prompts, and auxiliary functions. 55 00:03:44,619 --> 00:03:49,699 It works well for basic use cases, but refer to the LangGraph documentation for more advanced 56 00:03:49,699 --> 00:03:52,179 implementations. 57 00:03:52,179 --> 00:03:57,699 Define a search tool by wrapping a TavilySearchResults object with the @ tool decorator. 58 00:03:57,699 --> 00:04:02,699 Create a recommend_clothing tool that analyzes keywords in the LLM's response, 59 00:04:02,699 --> 00:04:09,179 such as rain or wet, to suggest appropriate clothing, as described in the docstring. 60 00:04:09,179 --> 00:04:14,380 The schema is essentially a dictionary-like state where the key is messages and the value 61 00:04:14,380 --> 00:04:18,940 is a growing list of messages exchanged during the conversation. 62 00:04:18,940 --> 00:04:24,500 Sequence[BaseMessage] holds a list of any message type, HumanMessage, AIMessage, or 63 00:04:24,500 --> 00:04:26,339 ToolMessage. 64 00:04:26,339 --> 00:04:30,500 Add_messages appends new messages each time a node runs. 65 00:04:30,500 --> 00:04:35,179 Annotated is required to turn the messages field into the correct form so that LangGraph 66 00:04:35,179 --> 00:04:38,239 knows to apply add_messages. 67 00:04:38,239 --> 00:04:43,660 The agent state object behaves like a dictionary where messages is a key and its value is a 68 00:04:43,660 --> 00:04:47,260 list that stores all the messages exchanged in the conversation. 69 00:04:47,260 --> 00:04:52,500 Next, load the GPT model using Lengchain's OpenAI wrapper. 70 00:04:52,500 --> 00:04:59,059 Then create a list of tools and a dictionary mapping each tool's name to the tool itself. 71 00:04:59,059 --> 00:05:03,540 Define a system message that instructs the agent to respond step-by-step and use tools 72 00:05:03,540 --> 00:05:05,299 when necessary. 73 00:05:05,299 --> 00:05:10,820 The agent_scratchpad variable automatically stores the reasoning history in the format. 74 00:05:10,820 --> 00:05:18,279 Thought, to action, to action input, to observation, along with the user's input sequence. 75 00:05:18,279 --> 00:05:24,200 While many implementations store the user's input, or question, and tool messages separately, 76 00:05:24,200 --> 00:05:27,880 for brevity, store everything in the scratchpad. 77 00:05:27,880 --> 00:05:34,000 Next, chain the prompt to the model and bind the tools to form the agent. 78 00:05:34,000 --> 00:05:37,880 Define the first node in the graph, which takes the agent state. 79 00:05:37,880 --> 00:05:41,920 The state is passed to the model through scratch_pad. 80 00:05:41,920 --> 00:05:47,320 The model's response is returned and appended to the message history. 81 00:05:47,320 --> 00:05:49,959 Create a tool calling node. 82 00:05:49,959 --> 00:05:53,079 Create a new graph using agent state to track the data. 83 00:05:53,079 --> 00:05:57,399 Add a node named agent that runs the call_model function. 84 00:05:57,399 --> 00:06:02,559 Next, add a node named tools that runs the tool_node function. 85 00:06:02,559 --> 00:06:07,559 Finally, add an edge from agent to tools. 86 00:06:07,559 --> 00:06:11,880 Define a should_continue function to control the agent's flow. 87 00:06:11,880 --> 00:06:14,859 Pair it with a conditional edge that decides the next step. 88 00:06:14,859 --> 00:06:18,720 It takes the state from the agent and retrieves the last message. 89 00:06:18,720 --> 00:06:24,160 Based on that message, the graph routes to either the tools node or the end node. 90 00:06:24,160 --> 00:06:28,119 If the message has no tool calls, it returns end. 91 00:06:28,119 --> 00:06:30,600 This maps to the end node in the graph. 92 00:06:30,600 --> 00:06:36,279 Otherwise, it returns continue, and conditional edge maps continue to the tools node. 93 00:06:36,279 --> 00:06:39,600 Here, agent is set at the starting node of the graph. 94 00:06:39,600 --> 00:06:43,640 It compiles the workflow into an executable graph. 95 00:06:43,640 --> 00:06:47,239 This function will help print out results from the graph. 96 00:06:47,239 --> 00:06:50,200 Define the input message from the user to start the graph. 97 00:06:50,200 --> 00:06:51,799 What's the weather like in Zurich? 98 00:06:51,799 --> 00:06:53,880 And what should I wear based on the temperature? 99 00:06:53,880 --> 00:06:59,760 Next, run the graph and print each state update step-by-step using the print_stream 100 00:06:59,760 --> 00:07:01,519 function. 101 00:07:01,519 --> 00:07:06,920 The trace of the full reasoning process starting from the human message can be seen. 102 00:07:06,920 --> 00:07:11,440 The first AI message includes a tool call to the search tool with its parameters. 103 00:07:11,440 --> 00:07:13,200 The tool returns a result. 104 00:07:13,200 --> 00:07:16,679 The second AI message calls the clothing recommendation tool. 105 00:07:16,679 --> 00:07:19,279 The second tool result is added to the state. 106 00:07:19,279 --> 00:07:24,640 Finally, the LLM generates the last AI message with the complete answer. 107 00:07:24,640 --> 00:07:26,760 You can sketch what's happening in the graph. 108 00:07:26,760 --> 00:07:31,640 The initial human message is passed to the call_model via the state. 109 00:07:31,640 --> 00:07:39,320 The LLM responds, and an AI message is appended to the state messages variable. 110 00:07:39,320 --> 00:07:43,160 The state is then passed to the should_continue node. 111 00:07:43,160 --> 00:07:47,519 The last message, an AI message, is extracted from the state. 112 00:07:47,519 --> 00:07:51,760 If it includes a tool call, the conditional edge returns continue and routes to the tool 113 00:07:51,760 --> 00:07:53,239 node. 114 00:07:53,239 --> 00:07:57,480 Similarly, the tool node extracts the tool call from the state. 115 00:07:57,480 --> 00:07:59,119 The tool name is identified. 116 00:07:59,119 --> 00:08:02,519 The tool parameters are selected based on the chosen tool. 117 00:08:02,519 --> 00:08:06,519 The result is wrapped in a tool message and passed back to the model. 118 00:08:06,519 --> 00:08:09,640 This process repeats until no more tool calls are made. 119 00:08:09,640 --> 00:08:12,959 At that point, the final response is generated. 120 00:08:12,959 --> 00:08:15,600 The graph reaches the end node. 121 00:08:15,600 --> 00:08:18,000 In this video, you learned that 122 00:08:18,000 --> 00:08:23,079 React agents perform step-by-step reasoning and use tools to answer complex queries. 123 00:08:23,079 --> 00:08:25,440 Their output follows a structured format. 124 00:08:25,440 --> 00:08:31,119 Thought ➝ action ➝ action input ➝ observation ➝ final answer. 125 00:08:31,119 --> 00:08:36,200 Tool results (observations) feed back into the reasoning process to guide next steps. 126 00:08:36,200 --> 00:08:41,359 LangGraph is used to implement the React workflow connecting reasoning and tool nodes. 127 00:08:41,359 --> 00:08:46,280 The process continues until a final answer is generated with no further tool calls required.