#Middleware

Middleware is LangChain 1.0's main extension point: hook into fixed stages of the agent loop so you don't rewrite the loop for every production concern.

#Why middleware?

Common production needs:

• Summarization (context limits)
• Long-term memory injection
• Human-in-the-loop before sensitive tools
• Subagent delegation
• Dynamic prompts / model routing
• Step and tool limits

Hooks sit around model and tool nodes; multiple middleware pieces compose (onion model).

#Hook types (conceptual)

Order: forward on the way in, reverse on the way out.

#Example: step limit (sketch)

from langchain.agents import create_agent
from langchain.agents.middleware import AgentMiddleware  # see official docs

class StepLimitMiddleware(AgentMiddleware):
    def __init__(self, max_steps: int = 10):
        self.max_steps = max_steps

    def before_model(self, state, runtime):
        ...

agent = create_agent(
    "openai:gpt-4.1-mini",
    tools=[...],
    middleware=[StepLimitMiddleware(max_steps=15)],
)

Exact base classes and signatures depend on your installed version—see official middleware docs.

#Built-in & community middleware

Examples: human-in-the-loop, summarization, model fallback, PII filters, prompt caching. Check LangChain blog and GitHub for community lists.

#Middleware vs LangGraph nodes

Use middleware for single-agent enhancements; LangGraph for orchestration patterns.

#Subagents

Middleware can spawn subagents with isolated context and merge results—see Deep Agents / delegation middleware.

#Debugging

Use LangSmith state diffs; add one middleware at a time; avoid heavy sync IO in hooks.

#Next steps

• Memory & Checkpoints
• LangGraph Workflows
• Production & LangSmith