LangChain create_agent() Function
create_agent() is the core function of LangChain. It creates a complete Agent graph (StateGraph), including all logic such as model invocation, tool execution, and loop control.
Syntax
The syntax of the create_agent() function is as follows:
from langchain.agents import create_agent
agent = create_agent(
model, # str | BaseChatModel：语言模型
tools=None, # Sequence：工具列表
*,
system_prompt=None, # str | SystemMessage：系统提示
middleware=(), # Sequence[AgentMiddleware]：中间件列表
response_format=None, # ResponseFormat | type：结构化输出配置
state_schema=None, # type[AgentState]：自定义状态结构
context_schema=None, # type：运行时上下文结构
checkpointer=None, # Checkpointer：对话持久化
store=None, # BaseStore：跨会话存储
interrupt_before=None, # list[str]：在哪些节点前暂停
interrupt_after=None, # list[str]：在哪些节点后暂停
debug=False, # bool：是否输出详细日志
name=None, # str：Agent 名称
cache=None, # BaseCache：缓存配置
)
model Parameter - Model Configuration
The model parameter accepts two forms: a string (processed by init_chat_model()) or an already constructed BaseChatModel instance.
Example
from langchain.chat_models import init_chat_model
# Method 1: Pass a string (most common)
# create_agent internally calls init_chat_model() for processing
agent = create_agent(
model="deepseek:deepseek-v4-flash",
system_prompt="You are the assistant of EXAMPLE",
)
# Method 2: Pass an already constructed model instance
# Suitable for scenarios where you need fine-grained control over model parameters
model = init_chat_model("deepseek:deepseek-v4-flash", temperature=0.3, max_tokens=500)
agent = create_agent(
model=model,
system_prompt="You are the assistant of EXAMPLE",
)
# Method 3: Pass a model instance with tools already bound
# Less common; usually let create_agent handle tool binding itself
model_with_tools = init_chat_model("deepseek:deepseek-v4-flash").bind_tools([...])
Method 1 (passing a string) is recommended. create_agent() automatically handles model initialization, tool binding, structured output, and other logic internally. Method 2 is suitable for scenarios where you need to use the same model instance outside the Agent.
tools Parameter - Tool List
The tools parameter accepts tools in three formats:
Example
from langchain.agents import create_agent
# Format 1: A function decorated with @tool (most common)
@tool
def search_course(keyword: str) -> str:
"""Search EXAMPLE courses"""
return f"Search results: {keyword} related courses"
# Format 2: A Pydantic BaseModel class
from pydantic import BaseModel, Field
class WeatherQuery(BaseModel):
"""Query weather"""
city: str = Field(description="City name")
# Format 3: A dictionary (describing remote tools or built-in tools)
mcp_tool = {
"type": "mcp",
"server_label": "weather_server",
"server_url": "https://weather.example.com/sse",
"allowed_tools": ["get_forecast"],
}
# Mixed usage
agent = create_agent(
model="deepseek:deepseek-v4-flash",
tools=[search_course, WeatherQuery, mcp_tool],
)
Passing None or an empty list means the Agent has no tools available; in this case, it is simply a pure conversation model:
Example
from langchain.messages import HumanMessage
# Agent without tools - equivalent to directly calling the model
agent = create_agent(
model="deepseek:deepseek-v4-flash",
tools=None,
system_prompt="You are the assistant of EXAMPLE",
)
result = agent.invoke({
"messages": [HumanMessage(content="Is Python suitable for complete beginners?")]
})
print(result["messages"][-1].content)
system_prompt Parameter - System Prompt
Defines the Agent's behavioral role and constraint rules. Supports strings and SystemMessage objects.
Example
from langchain.messages import SystemMessage
# Method 1: String (simple and direct)
agent = create_agent(
model="deepseek:deepseek-v4-flash",
system_prompt="You are the learning consultant of EXAMPLE. Keep answers concise, no more than 100 characters.",
)
# Method 2: SystemMessage object (can be reused across multiple Agents)
system_msg = SystemMessage(
content="You are the learning consultant of EXAMPLE. Keep answers concise, no more than 100 characters."
)
agent = create_agent(
model="deepseek:deepseek-v4-flash",
system_prompt=system_msg,
)
system_prompt is optional, but if not provided, the model will respond in the role of a "general-purpose assistant". For applications with clear business scenarios, it is recommended to always set system_prompt to constrain the model's behavior boundaries.
state_schema Parameter - Custom State
The default AgentState only includes messages, jump_to, and structured_response. If you need additional state fields, you can extend it:
Example
from langchain.agents import create_agent, AgentState
from langchain.messages import HumanMessage
from langchain.tools import tool, InjectedState
from typing_extensions import TypedDict
# Extend AgentState and add custom fields
class LearningAgentState(AgentState):
"""Custom state, adding fields related to learning progress"""
user_level: str # User level
completed_topics: list[str] # List of completed topics
@tool
def track_progress(
topic: str,
state: Annotated[dict, InjectedState],
) -> str:
"""Record the user's learning progress.
Args:
topic: The name of the topic just completed
"""
completed = state.get("completed_topics", [])
completed.append(topic)
return (
f"Learning progress recorded. Currently completed {len(completed)} topics:"
f"{', '.join(completed)}"
)
agent = create_agent(
model="deepseek:deepseek-v4-flash",
tools=[track_progress],
state_schema=LearningAgentState, # Using custom state
system_prompt="You are the learning assistant of EXAMPLE.",
)
# The initial value of the custom state must be provided at runtime
result = agent.invoke({
"messages": [HumanMessage(content="I have finished learning Python basics. Please record it for me.")],
"user_level": "Beginner",
"completed_topics": ["HTML Basics"],
})
print(f"User level: {result.get('user_level')}")
print(f"Completed topics: {result.get('completed_topics')}")
print(f"Reply: {result['messages'][-1].content[:100]}")
Output:
User level: 入门 Completed topics: ['HTML 基础', 'Python 基础'] Reply: Learning progress recorded. Completed 2 topics: HTML Basics, Python Basics
Return Value - CompiledStateGraph
create_agent() returns aCompiledStateGraphobject, which is the compiled graph of LangGraph and provides multiple ways to run:
| Method | Description | Use Case |
|---|---|---|
| invoke(input, config) | Synchronous run, wait for the complete result | Scripts, simple interfaces |
| ainvoke(input, config) | Asynchronous run, wait for the complete result | Web services |
| stream(input, config, stream_mode) | Synchronous streaming run | Display intermediate steps in real time |
| astream(input, config, stream_mode) | Asynchronous streaming run | WebSocket、SSE |
| get_state(config) | Get the current state | View/restore conversation state |
| update_state(config, values) | Update state | Manually modify conversation state |