ReAct Agents: Synergizing Reasoning and Acting

ReAct (Reasoning + Acting) is a paradigm that combines chain-of-thought reasoning with action-taking, enabling language models to solve complex tasks by interleaving thought and interaction with external tools.

The Core Insight

Traditional approaches separate reasoning from acting:

Approach	Limitation
Chain-of-Thought only	Can't interact with world, facts may be stale
Action only	No explicit reasoning, hard to debug
ReAct	Best of both—reason about actions, act to gather information

The ReAct Loop

Question: What is the elevation of the highest mountain in the country 
          where the 2024 Olympics were held?

Thought 1: I need to find where the 2024 Olympics were held.
Action 1: Search[2024 Olympics location]
Observation 1: The 2024 Summer Olympics were held in Paris, France.

Thought 2: The country is France. Now I need to find the highest mountain in France.
Action 2: Search[highest mountain France]
Observation 2: Mont Blanc is the highest mountain in France at 4,808 meters.

Thought 3: I have the answer. Mont Blanc's elevation is 4,808 meters.
Action 3: Finish[4,808 meters]

Implementation

Basic ReAct Agent

class ReActAgent:
    def __init__(self, llm, tools):
        self.llm = llm
        self.tools = {tool.name: tool for tool in tools}
        self.max_steps = 10
    
    def run(self, question: str) -> str:
        prompt = self.build_initial_prompt(question)
        
        for step in range(self.max_steps):
            # Generate thought and action
            response = self.llm.generate(prompt)
            
            # Parse the response
            thought, action, action_input = self.parse_response(response)
            
            if action == "Finish":
                return action_input
            
            # Execute action
            observation = self.tools[action].run(action_input)
            
            # Update prompt with observation
            prompt += f"\nThought {step+1}: {thought}"
            prompt += f"\nAction {step+1}: {action}[{action_input}]"
            prompt += f"\nObservation {step+1}: {observation}"
        
        return "Max steps reached without answer"
    
    def parse_response(self, response: str):
        # Extract Thought, Action, and Action Input
        thought_match = re.search(r"Thought:(.+?)Action:", response, re.DOTALL)
        action_match = re.search(r"Action:(.+?)\[(.+?)\]", response)
        
        thought = thought_match.group(1).strip() if thought_match else ""
        action = action_match.group(1).strip() if action_match else ""
        action_input = action_match.group(2).strip() if action_match else ""
        
        return thought, action, action_input

Prompt Template

REACT_PROMPT = """Answer the following question by reasoning step-by-step 
and using tools when needed.

Available tools:
- Search[query]: Search the web for information
- Lookup[term]: Look up a term in the current context
- Calculate[expression]: Perform mathematical calculations
- Finish[answer]: Return the final answer

Question: {question}

Think step by step. For each step:
1. Write "Thought:" followed by your reasoning
2. Write "Action:" followed by the tool and input
3. Wait for "Observation:" with the result
4. Repeat until you can provide a final answer

Begin!
"""

ReAct vs Other Approaches

Chain-of-Thought (CoT)

CoT: Think → Think → Think → Answer
     (no external information)

ReAct: Think → Act → Observe → Think → Act → Observe → Answer
       (grounded in real data)

Action-Only

Action-Only: Act → Act → Act → Answer
             (no explicit reasoning, brittle)

ReAct: Think → Act → Think → Act → Answer
       (reasoning explains and guides actions)

Advanced Patterns

ReAct with Self-Reflection

class ReflectiveReActAgent(ReActAgent):
    def reflect(self, trajectory: List[Step]) -> str:
        reflection_prompt = f"""
        Review this problem-solving trajectory:
        {format_trajectory(trajectory)}
        
        What went well? What could be improved?
        Should we try a different approach?
        """
        return self.llm.generate(reflection_prompt)
    
    def run_with_reflection(self, question: str) -> str:
        trajectory = []
        
        for attempt in range(3):
            result, steps = self.run_and_record(question)
            trajectory.extend(steps)
            
            if self.is_confident(result):
                return result
            
            reflection = self.reflect(trajectory)
            # Use reflection to adjust strategy

Hierarchical ReAct

High-Level Agent
├── Thought: Break down into subtasks
├── Action: Delegate to SubAgent1
│   └── SubAgent1 runs full ReAct loop
├── Observation: SubAgent1 result
├── Action: Delegate to SubAgent2
│   └── SubAgent2 runs full ReAct loop
├── Observation: SubAgent2 result
└── Finish: Combine results

ReAct with Memory

class MemoryReActAgent(ReActAgent):
    def __init__(self, llm, tools, memory_store):
        super().__init__(llm, tools)
        self.memory = memory_store
    
    def run(self, question: str) -> str:
        # Retrieve relevant past experiences
        relevant_memories = self.memory.search(question, k=3)
        
        # Include in prompt
        prompt = self.build_prompt_with_memories(question, relevant_memories)
        
        result = super().run(question)
        
        # Store this trajectory for future reference
        self.memory.store(question, self.trajectory, result)
        
        return result

Tool Design for ReAct

Good Tool Design

class SearchTool:
    name = "Search"
    description = "Search the web. Input: search query string"
    
    def run(self, query: str) -> str:
        results = web_search(query)
        # Return concise, actionable information
        return self.format_results(results, max_length=500)

Tool Selection

TOOL_SELECTION_PROMPT = """
Given the current thought, select the most appropriate tool:

Thought: {thought}

Available tools:
1. Search - for finding factual information
2. Calculate - for mathematical operations
3. Code - for running Python code
4. Lookup - for searching in current context

Which tool should be used? Explain briefly, then select.
"""

Evaluation

Metrics

Metric	Description
Task Success Rate	% of tasks completed correctly
Efficiency	Average number of steps to complete
Reasoning Quality	Are thoughts logical and helpful?
Tool Use Accuracy	Are tools used appropriately?
Hallucination Rate	% of false claims in reasoning

Benchmarks

HotpotQA: Multi-hop question answering
FEVER: Fact verification
ALFWorld: Embodied tasks in text environments
WebShop: Web navigation and shopping

Limitations

Error propagation: Wrong early steps compound
Verbosity: Lots of tokens for reasoning
Tool dependency: Limited by available tools
Prompt sensitivity: Performance varies with prompt design

Integration with Modern Frameworks

LangChain

from langchain.agents import create_react_agent
from langchain.tools import Tool

tools = [
    Tool(name="Search", func=search_func, description="..."),
    Tool(name="Calculator", func=calc_func, description="...")
]

agent = create_react_agent(llm, tools, prompt)
result = agent.invoke({"input": question})

LlamaIndex

from llama_index.agent import ReActAgent
from llama_index.tools import QueryEngineTool

agent = ReActAgent.from_tools(
    tools=[query_engine_tool],
    llm=llm,
    verbose=True
)

References

ReAct shows that the combination of reasoning and acting is more powerful than either alone—a principle that extends far beyond AI to how humans solve problems.