langchain-middleware

To add human-in-the-loop approval for LangChain agents, implement middleware that pauses execution before dangerous tool calls. This requires a checkpointer to persist state and uses hooks like wrap_tool_call to intercept actions and wait for human decisions.

LangChain middleware hooks are interception points that manage agent execution flows. Specific hooks include before_model, after_model, wrap_tool_call, before_agent, and after_agent, enabling custom policies for error handling, logging, and pausing dangerous actions.

LangChain human-in-the-loop middleware supports resuming execution after a human decision using command resume. This allows workflows to continue execution seamlessly whether the human operator chooses to approve, edit, or reject the paused tool call.

You can configure per-tool interruption policies for LangChain agents. This enables customizable workflows where only specific dangerous tools require human approval, while safe tools execute automatically without pausing.

A checkpointer is required to use human-in-the-loop middleware with LangChain. The checkpointer persists agent state, allowing the workflow to pause before a dangerous tool call and resume execution later after human review.

To handle tool call errors and retries using LangChain middleware, implement custom hooks like wrap_tool_call. These middleware hooks intercept tool calls, allowing you to inject error handling logic, log actions, and automatically retry failed operations.