ReAct & Tool Use
LLMs are powerful reasoners, but they can't take actions on their own. Tool use gives LLMs the ability to call functions — search the web, run code, query databases — and then reason about the results. This turns a chatbot into an agent.
What is an Agent?
A traditional LLM responds to a prompt with text. An agent:
- Receives prompt
- Generates text response
- Done
- Receives goal
- Reasons about what to do
- Calls a tool (web search, code, API)
- Observes result
- Repeats until goal is achieved
The ReAct Pattern
ReAct (Reasoning + Acting) is the foundational pattern for LLM agents. The model iterates through: Thought → Action → Observation until it has enough information to give a final answer.
I need to find the current weather in Tokyo. I'll use the weather tool.
get_weather(location="Tokyo, Japan")
"Tokyo: 22°C, partly cloudy, humidity 65%"
I have the weather. I can now answer the user's question.
The current weather in Tokyo is 22°C and partly cloudy.
Function Calling and Tool Use
Modern LLM APIs support "function calling" — you define tools in JSON schema format, and the model decides when and how to call them.
import anthropic
client = anthropic.Anthropic()
# Define tools
tools = [
{
"name": "get_weather",
"description": "Get current weather for a city",
"input_schema": {
"type": "object",
"properties": {
"location": {"type": "string", "description": "City name, e.g. 'Tokyo, Japan'"}
},
"required": ["location"]
}
},
{
"name": "run_python",
"description": "Execute Python code and return the output",
"input_schema": {
"type": "object",
"properties": {
"code": {"type": "string", "description": "Python code to execute"}
},
"required": ["code"]
}
}
]
response = client.messages.create(
model="claude-sonnet-4-6",
max_tokens=1024,
tools=tools,
messages=[{"role": "user", "content": "What's 2^50 + the current year?"}]
)
# Check if model wants to call a tool
if response.stop_reason == "tool_use":
tool_call = next(b for b in response.content if b.type == "tool_use")
print(f"Model wants to call: {tool_call.name}")
print(f"With inputs: {tool_call.input}")Building a Complete Agent Loop
import anthropic
import subprocess
client = anthropic.Anthropic()
def run_python(code: str) -> str:
result = subprocess.run(['python', '-c', code], capture_output=True, text=True, timeout=10)
return result.stdout or result.stderr
def agent_loop(user_message: str, max_iterations: int = 10) -> str:
messages = [{"role": "user", "content": user_message}]
for _ in range(max_iterations):
response = client.messages.create(
model="claude-sonnet-4-6",
max_tokens=2048,
tools=[{
"name": "run_python",
"description": "Run Python code",
"input_schema": {"type": "object", "properties": {"code": {"type": "string"}}, "required": ["code"]}
}],
messages=messages
)
# If done, return the text response
if response.stop_reason == "end_turn":
return response.content[0].text
# Otherwise, execute the tool and continue
tool_use = next(b for b in response.content if b.type == "tool_use")
result = run_python(tool_use.input['code'])
messages.append({"role": "assistant", "content": response.content})
messages.append({
"role": "user",
"content": [{"type": "tool_result", "tool_use_id": tool_use.id, "content": result}]
})
return "Max iterations reached"
print(agent_loop("What is the sum of all prime numbers under 100?"))Common Agent Tools
Tavily, Serper, Google Search API. Get real-time information the model wasn't trained on.
Python REPL, Jupyter kernel. Let the model calculate, analyse data, create charts.
Read/write files, search directories. Build file management agents.
Call any REST API. Weather, CRM, databases, internal services.
Natural language to SQL. The model writes queries and interprets results.
Send emails, Slack messages, create calendar events via tool calls.
Frequently Asked Questions
How do I prevent agents from taking harmful actions?
Always implement a "human-in-the-loop" for irreversible actions (sending emails, deleting files, making purchases). Use tool descriptions to constrain scope. Add a confirmation step before executing high-risk actions. Run tools in sandboxed environments (Docker containers) to prevent system access.
What is the difference between ReAct and chain-of-thought?
Chain-of-thought (CoT) is pure reasoning — the model thinks step by step but doesn't take external actions. ReAct extends CoT with action steps: the model can call real tools and incorporate the results into its reasoning. ReAct agents can access information beyond their training data.