Langchain

# LangChain Agents Guide

Complete guide to building agents with ReAct, tool calling, and streaming.

## What are agents?

Agents combine language models with tools to solve complex tasks through reasoning and action:

1. **Reasoning**: LLM decides what to do
2. **Acting**: Execute tools based on reasoning
3. **Observation**: Receive tool results
4. **Loop**: Repeat until task complete

This is the **ReAct pattern** (Reasoning + Acting).

## Basic agent creation

```python
from langchain.agents import create_agent
from langchain_anthropic import ChatAnthropic

# Define tools
def calculator(expression: str) -> str:
    """Evaluate a math expression."""
    return str(eval(expression))

def search(query: str) -> str:
    """Search for information."""
    return f"Results for: {query}"

# Create agent
agent = create_agent(
    model=ChatAnthropic(model="claude-sonnet-4-5-20250929"),
    tools=[calculator, search],
    system_prompt="You are a helpful assistant. Use tools when needed."
)

# Run agent
result = agent.invoke({
    "messages": [{"role": "user", "content": "What is 25 * 17?"}]
})
print(result["messages"][-1].content)
```

## Agent components

### 1. Model - The reasoning engine

```python
from langchain_openai import ChatOpenAI
from langchain_anthropic import ChatAnthropic

# OpenAI
model = ChatOpenAI(model="gpt-4o", temperature=0)

# Anthropic (better for complex reasoning)
model = ChatAnthropic(model="claude-sonnet-4-5-20250929", temperature=0)

# Dynamic model selection
def select_model(task_complexity: str):
    if task_complexity == "high":
        return ChatAnthropic(model="claude-sonnet-4-5-20250929")
    else:
        return ChatOpenAI(model="gpt-4o-mini")
```

### 2. Tools - Actions the agent can take

```python
from langchain.tools import tool

# Simple function tool
@tool
def get_current_time() -> str:
    """Get the current time."""
    from datetime import datetime
    return datetime.now().strftime("%H:%M:%S")

# Tool with parameters
@tool
def fetch_weather(city: str, units: str = "fahrenheit") -> str:
    """Fetch weather for a city.

    Args:
        city: City name
        units: Temperature units (fahrenheit or celsius)
    """
    # Your weather API call here
    return f"Weather in {city}: 72°{units[0].upper()}"

# Tool with error handling
@tool
def risky_api_call(endpoint: str) -> str:
    """Call an external API that might fail."""
    try:
        response = requests.get(endpoint, timeout=5)
        return response.text
    except Exception as e:
        return f"Error calling API: {str(e)}"
```

### 3. System prompt - Agent behavior

```python
# General assistant
system_prompt = "You are a helpful assistant. Use tools when needed."

# Domain expert
system_prompt = """You are a financial analyst assistant.
- Use the calculator for precise calculations
- Search for recent financial data
- Provide data-driven recommendations
- Always cite your sources"""

# Constrained agent
system_prompt = """You are a customer support agent.
- Only use search_kb tool to find answers
- If answer not found, escalate to human
- Be concise and professional
- Never make up information"""
```

## Agent types

### 1. Tool-calling agent (recommended)

Uses native function calling for best performance:

```python
from langchain.agents import create_tool_calling_agent, AgentExecutor
from langchain.prompts import ChatPromptTemplate

# Create prompt
prompt = ChatPromptTemplate.from_messages([
    ("system", "You are a helpful assistant"),
    ("human", "{input}"),
    ("placeholder", "{agent_scratchpad}"),
])

# Create agent
agent = create_tool_calling_agent(
    llm=model,
    tools=[calculator, search],
    prompt=prompt
)

# Wrap in executor
agent_executor = AgentExecutor(
    agent=agent,
    tools=[calculator, search],
    verbose=True,
    max_iterations=5,
    handle_parsing_errors=True
)

# Run
result = agent_executor.invoke({"input": "What is the weather in Paris?"})
```

### 2. ReAct agent (reasoning trace)

Show