Middleware

Middleware hooks into an agent before, after, or around model and tool calls. Use it for cross-cutting behavior such as logging, limits, redaction, retries, fallbacks, and human approval.

Tiny example

from langchain.agents import create_agent
from langchain.agents.middleware import ToolCallLimitMiddleware

agent = create_agent(
    model="openai:gpt-5.4-mini",
    tools=[search_orders],
    middleware=[
        ToolCallLimitMiddleware(tool_name="search_orders", run_limit=3),
    ],
)

Hook mental model

Hook style	Use it for
before_*	Validate or add state before a step
after_*	Inspect a result or record state
wrap_*	Control whether the wrapped call runs zero, one, or several times

Middleware hooks run inside the compiled LangGraph returned by create_agent(); middleware is not a second runtime.

Failure note

Order matters when middleware wraps other middleware. Retrying outside a handler that performs a side effect can repeat that side effect. Test the actual order and keep authorization in the tool or service that owns the data.

Related

create_agent · Retries and backoff · Callbacks and tracing