01. ReAct Pattern from Scratch

Overview

In this session, we implement the ReAct (Reasoning + Acting) pattern from scratch using pure Python—no frameworks, no magic. This helps you understand exactly how AI agents work under the hood.

What is ReAct?

ReAct is a pattern where an LLM alternates between:

1. Thought - Reasoning about what to do next
2. Action - Calling a tool to get information
3. Observation - Receiving and processing results

Key Concepts

The Agent Loop

# Pseudocode for ReAct loop
while not done:
    response = llm.generate(messages)
 
    if "Final Answer" in response:
        return parse_final_answer(response)
 
    action = parse_action(response)
    observation = execute_tool(action)
 
    messages.append(observation)

System Prompt Design

The system prompt is crucial—it tells the LLM exactly how to format its responses:

You are an AI agent that can use tools.

Available tools:
1. calculate[expression] - Performs math calculations
2. search[query] - Searches for information

Format your response as:
Thought: [your reasoning]
Action: [tool_name][argument]

Regex Parsing

We use regular expressions to extract structured information from the LLM's free-form text output:

import re
 
def parse_action(text):
    pattern = r"Action:\s*(\w+)\[(.+?)\]"
    match = re.search(pattern, text)
    if match:
        return match.group(1), match.group(2)
    return None

Hands-on Practice

In the notebook, you will:

1. Define Tools - Create calculate and search functions
2. Design System Prompt - Craft instructions for the LLM
3. Implement Parser - Extract actions from LLM output
4. Build Agent Class - Put it all together in ReActAgent
5. Test & Debug - Run various queries and analyze behavior

Limitations of This Approach

Next Steps

In Session 2: Tool Calling, we'll solve these issues using:

• Pydantic for type-safe schemas
• OpenAI Function Calling for structured output
• Proper error handling