langchain-fundamentals
Construisez des agents LangChain de production avec create_agent(), des outils et des modèles de middleware. Utilisez create_agent() avec un modèle, une liste d'outils et un prompt système ; configurez la persistance d'état avec un checkpointer et un thread_id pour la mémoire de conversation entre les invocations. Définissez des outils via le décorateur @tool (Python) ou la fonction tool() (TypeScript) avec des descriptions claires pour que les agents sachent quand les appeler. Ajoutez du middleware comme HumanInTheLoopMiddleware pour les workflows d'approbation, la gestion personnalisée des erreurs et l'intervention humaine...
npx skills add https://github.com/langchain-ai/langchain-skills --skill langchain-fundamentals<create_agent>
Creating Agents with create_agent
create_agent() is the recommended way to build agents. It handles the agent loop, tool execution, and state management.
Agent Configuration Options
| Parameter | Purpose | Example |
|---|---|---|
model | LLM to use | "anthropic:claude-sonnet-4-5" or model instance |
tools | List of tools | [search, calculator] |
system_prompt / systemPrompt | Agent instructions | "You are a helpful assistant" |
checkpointer | State persistence | MemorySaver() |
middleware | Processing hooks | [HumanInTheLoopMiddleware] (Python) / [humanInTheLoopMiddleware({...})] (TypeScript) |
| </create_agent> |
from langchain.agents import create_agent
from langchain_core.tools import tool
@tool
def get_weather(location: str) -> str:
"""Get current weather for a location.
Args:
location: City name
"""
return f"Weather in {location}: Sunny, 72F"
agent = create_agent(
model="anthropic:claude-sonnet-4-5",
tools=[get_weather],
system_prompt="You are a helpful assistant."
)
result = agent.invoke({
"messages": [{"role": "user", "content": "What's the weather in Paris?"}]
})
print(result["messages"][-1].content)
import { createAgent } from "langchain";
import { tool } from "@langchain/core/tools";
import { z } from "zod";
const getWeather = tool(
async ({ location }) => `Weather in ${location}: Sunny, 72F`,
{
name: "get_weather",
description: "Get current weather for a location.",
schema: z.object({ location: z.string().describe("City name") }),
}
);
const agent = createAgent({
model: "anthropic:claude-sonnet-4-5",
tools: [getWeather],
systemPrompt: "You are a helpful assistant.",
});
const result = await agent.invoke({
messages: [{ role: "user", content: "What's the weather in Paris?" }],
});
console.log(result.messages[result.messages.length - 1].content);
Add MemorySaver checkpointer to maintain conversation state across invocations.
from langchain.agents import create_agent
from langgraph.checkpoint.memory import MemorySaver
checkpointer = MemorySaver()
agent = create_agent(
model="anthropic:claude-sonnet-4-5",
tools=[search],
checkpointer=checkpointer,
)
config = {"configurable": {"thread_id": "user-123"}}
agent.invoke({"messages": [{"role": "user", "content": "My name is Alice"}]}, config=config)
result = agent.invoke({"messages": [{"role": "user", "content": "What's my name?"}]}, config=config)
# Agent remembers: "Your name is Alice"
Add MemorySaver checkpointer to maintain conversation state across invocations.
import { createAgent } from "langchain";
import { MemorySaver } from "@langchain/langgraph";
const checkpointer = new MemorySaver();
const agent = createAgent({
model: "anthropic:claude-sonnet-4-5",
tools: [search],
checkpointer,
});
const config = { configurable: { thread_id: "user-123" } };
await agent.invoke({ messages: [{ role: "user", content: "My name is Alice" }] }, config);
const result = await agent.invoke({ messages: [{ role: "user", content: "What's my name?" }] }, config);
// Agent remembers: "Your name is Alice"
## Defining Tools
Tools are functions that agents can call. Use the @tool decorator (Python) or tool() function (TypeScript).
from langchain_core.tools import tool
@tool
def add(a: float, b: float) -> float:
"""Add two numbers.
Args:
a: First number
b: Second number
"""
return a + b
import { tool } from "@langchain/core/tools";
import { z } from "zod";
const add = tool(
async ({ a, b }) => a + b,
{
name: "add",
description: "Add two numbers.",
schema: z.object({
a: z.number().describe("First number"),
b: z.number().describe("Second number"),
}),
}
);
## Middleware for Agent Control
Middleware intercepts the agent loop to add human approval, error handling, logging, and more. A deep understanding of middleware is essential for production agents — use HumanInTheLoopMiddleware (Python) / humanInTheLoopMiddleware (TypeScript) for approval workflows, and @wrap_tool_call (Python) / createMiddleware (TypeScript) for custom hooks.
Key imports:
from langchain.agents.middleware import HumanInTheLoopMiddleware, wrap_tool_call
import { humanInTheLoopMiddleware, createMiddleware } from "langchain";
Key patterns:
- HITL:
middleware=[HumanInTheLoopMiddleware(interrupt_on={"dangerous_tool": True})]— requirescheckpointer+thread_id - Resume after interrupt:
agent.invoke(Command(resume={"decisions": [{"type": "approve"}]}), config=config) - Custom middleware:
@wrap_tool_calldecorator (Python) orcreateMiddleware({ wrapToolCall: ... })(TypeScript)
<structured_output>
Structured Output
Get typed, validated responses from agents using response_format or with_structured_output().
from langchain.agents import create_agent
from pydantic import BaseModel, Field
class ContactInfo(BaseModel):
name: str
email: str
phone: str = Field(description="Phone number with area code")
# Option 1: Agent with structured output
agent = create_agent(model="gpt-4.1", tools=[search], response_format=ContactInfo)
result = agent.invoke({"messages": [{"role": "user", "content": "Find contact for John"}]})
print(result["structured_response"]) # ContactInfo(name='John', ...)
# Option 2: Model-level structured output (no agent needed)
from langchain_openai import ChatOpenAI
model = ChatOpenAI(model="gpt-4.1")
structured_model = model.with_structured_output(ContactInfo)
response = structured_model.invoke("Extract: John, [email protected], 555-1234")
# ContactInfo(name='John', email='[email protected]', phone='555-1234')
import { ChatOpenAI } from "@langchain/openai";
import { z } from "zod";
const ContactInfo = z.object({
name: z.string(),
email: z.string().email(),
phone: z.string().describe("Phone number with area code"),
});
// Model-level structured output
const model = new ChatOpenAI({ model: "gpt-4.1" });
const structuredModel = model.withStructuredOutput(ContactInfo);
const response = await structuredModel.invoke("Extract: John, [email protected], 555-1234");
// { name: 'John', email: '[email protected]', phone: '555-1234' }
<model_config>
Model Configuration
create_agent accepts model strings ("anthropic:claude-sonnet-4-5", "openai:gpt-4.1") or model instances for custom settings:
from langchain_anthropic import ChatAnthropic
agent = create_agent(model=ChatAnthropic(model="claude-sonnet-4-5", temperature=0), tools=[...])
</model_config>
Clear descriptions help the agent know when to use each tool.# WRONG: Vague or missing description
@tool
def bad_tool(input: str) -> str:
"""Does stuff."""
return "result"
# CORRECT: Clear, specific description with Args
@tool
def search(query: str) -> str:
"""Search the web for current information about a topic.
Use this when you need recent data or facts.
Args:
query: The search query (2-10 words recommended)
"""
return web_search(query)
Clear descriptions help the agent know when to use each tool.
// WRONG: Vague description
const badTool = tool(async ({ input }) => "result", {
name: "bad_tool",
description: "Does stuff.", // Too vague!
schema: z.object({ input: z.string() }),
});
// CORRECT: Clear, specific description
const search = tool(async ({ query }) => webSearch(query), {
name: "search",
description: "Search the web for current information about a topic. Use this when you need recent data or facts.",
schema: z.object({
query: z.string().describe("The search query (2-10 words recommended)"),
}),
});
Add checkpointer and thread_id for conversation memory across invocations.
# WRONG: No persistence - agent forgets between calls
agent = create_agent(model="anthropic:claude-sonnet-4-5", tools=[search])
agent.invoke({"messages": [{"role": "user", "content": "I'm Bob"}]})
agent.invoke({"messages": [{"role": "user", "content": "What's my name?"}]})
# Agent doesn't remember!
# CORRECT: Add checkpointer and thread_id
from langgraph.checkpoint.memory import MemorySaver
agent = create_agent(
model="anthropic:claude-sonnet-4-5",
tools=[search],
checkpointer=MemorySaver(),
)
config = {"configurable": {"thread_id": "session-1"}}
agent.invoke({"messages": [{"role": "user", "content": "I'm Bob"}]}, config=config)
agent.invoke({"messages": [{"role": "user", "content": "What's my name?"}]}, config=config)
# Agent remembers: "Your name is Bob"
Add checkpointer and thread_id for conversation memory across invocations.
// WRONG: No persistence
const agent = createAgent({ model: "anthropic:claude-sonnet-4-5", tools: [search] });
await agent.invoke({ messages: [{ role: "user", content: "I'm Bob" }] });
await agent.invoke({ messages: [{ role: "user", content: "What's my name?" }] });
// Agent doesn't remember!
// CORRECT: Add checkpointer and thread_id
import { MemorySaver } from "@langchain/langgraph";
const agent = createAgent({
model: "anthropic:claude-sonnet-4-5",
tools: [search],
checkpointer: new MemorySaver(),
});
const config = { configurable: { thread_id: "session-1" } };
await agent.invoke({ messages: [{ role: "user", content: "I'm Bob" }] }, config);
await agent.invoke({ messages: [{ role: "user", content: "What's my name?" }] }, config);
// Agent remembers: "Your name is Bob"
Set recursion_limit in the invoke config to prevent runaway agent loops.
# WRONG: No iteration limit - could loop forever
result = agent.invoke({"messages": [("user", "Do research")]})
# CORRECT: Set recursion_limit in config
result = agent.invoke(
{"messages": [("user", "Do research")]},
config={"recursion_limit": 10}, # Stop after 10 steps
)
Set recursionLimit in the invoke config to prevent runaway agent loops.
// WRONG: No iteration limit
const result = await agent.invoke({ messages: [["user", "Do research"]] });
// CORRECT: Set recursionLimit in config
const result = await agent.invoke(
{ messages: [["user", "Do research"]] },
{ recursionLimit: 10 }, // Stop after 10 steps
);
Access the messages array from the result, not result.content directly.
# WRONG: Trying to access result.content directly
result = agent.invoke({"messages": [{"role": "user", "content": "Hello"}]})
print(result.content) # AttributeError!
# CORRECT: Access messages from result dict
result = agent.invoke({"messages": [{"role": "user", "content": "Hello"}]})
print(result["messages"][-1].content) # Last message content
Access the messages array from the result, not result.content directly.
// WRONG: Trying to access result.content directly
const result = await agent.invoke({ messages: [{ role: "user", content: "Hello" }] });
console.log(result.content); // undefined!
// CORRECT: Access messages from result object
const result = await agent.invoke({ messages: [{ role: "user", content: "Hello" }] });
console.log(result.messages[result.messages.length - 1].content); // Last message content