Kenhuang Academy
  • English

      • Models & Messages

      LangChain unifies messages and model interfaces via langchain-core. In 1.0, use provider:model strings and standardized content blocks.

      Model identifiers

      from langchain.agents import create_agent

agent = create_agent("openai:gpt-4.1-mini", tools=[])
agent = create_agent("anthropic:claude-sonnet-4-6", tools=[])
agent = create_agent("google_genai:gemini-2.0-flash", tools=[])

      Format: provider:model_id; integrations read API keys from the environment.

      Explicit model instances

      For temperature, base_url, proxies, etc.:

      from langchain_openai import ChatOpenAI
from langchain.agents import create_agent

model = ChatOpenAI(model="gpt-4.1-mini", temperature=0)
agent = create_agent(model, tools=[])

      Same pattern for Anthropic, Google, and others.

      Message types

      Agent messages typically include:

      TypeRole
      HumanMessageUser input
      AIMessageModel reply (may include tool_calls)
      ToolMessageTool result
      SystemMessageSystem instructions
      from langchain_core.messages import HumanMessage, AIMessage

result = agent.invoke({"messages": [HumanMessage(content="Hello")]})
last = result["messages"][-1]

      System prompt

      agent = create_agent(
    "openai:gpt-4.1-mini",
    tools=[],
    system_prompt="You are a concise technical assistant.",
)

      Use middleware for dynamic system prompts.

      Tool calling

      Models must support tool calling. Manual bind (outside create_agent):

      llm_with_tools = ChatOpenAI(model="gpt-4.1-mini").bind_tools([my_tool])
msg = llm_with_tools.invoke("What is 2 times 3?")

      create_agent handles bind + loop for you.

      Structured output

      Agent response_format:

      class Plan(BaseModel):
    steps: list[str]

agent = create_agent("openai:gpt-4.1-mini", tools=[], response_format=Plan)

      Model with_structured_output: for single-step nodes in workflows.

      Content blocks

      Rich/multimodal content is normalized as content blocks across providers.

      Dynamic model selection

      Middleware can swap models by task complexity for cost/quality tradeoffs.

      Next steps