langchain-middleware

1<overview>
2Middleware patterns for production LangChain agents:
3 
4- **HumanInTheLoopMiddleware** / **humanInTheLoopMiddleware**: Pause before dangerous tool calls for human approval
5- **Custom middleware**: Intercept tool calls for error handling, logging, retry logic
6- **Command resume**: Continue execution after human decisions (approve, edit, reject)
7 
8**Requirements:** Checkpointer + thread_id config for all HITL workflows.
9</overview>
10 
11---
12 
13## Human-in-the-Loop
14 
15<ex-basic-hitl-setup>
16<python>
17Set up an agent with HITL middleware that pauses before sending emails for approval.
18 
19```python
20from langchain.agents import create_agent
21from langchain.agents.middleware import HumanInTheLoopMiddleware
22from langgraph.checkpoint.memory import MemorySaver
23from langchain.tools import tool
24 
25@tool
26def send_email(to: str, subject: str, body: str) -> str:
27    """Send an email."""
28    return f"Email sent to {to}"
29 
30agent = create_agent(
31    model="gpt-4.1",
32    tools=[send_email],
33    checkpointer=MemorySaver(),  # Required for HITL
34    middleware=[
35        HumanInTheLoopMiddleware(
36            interrupt_on={
37                "send_email": {"allowed_decisions": ["approve", "edit", "reject"]},
38            }
39        )
40    ],
41)
42```
43</python>
44<typescript>
45Set up an agent with HITL that pauses before sending emails for human approval.
46 
47```typescript
48import { createAgent, humanInTheLoopMiddleware } from "langchain";
49import { MemorySaver } from "@langchain/langgraph";
50…