langchain-middleware

Human-in-the-loop approval in LangChain agents is enabled by integrating HumanInTheLoopMiddleware to pause tool calls, allowing humans to approve, edit, or reject actions before execution. This gates dangerous operations like emails or API calls.

LangChain agent workflows support custom middleware hooks including before_model, after_model, wrap_tool_call, before_agent, and after_agent. These hooks provide extensible control points to inject validation or custom logic throughout the agent pipeline.

LangChain supports resuming paused agent workflows using Command resume patterns. This allows workflows to continue execution after a human reviewer approves, edits, or rejects a pending tool action within a thread-scoped state.

You can configure per-tool human-in-the-loop policies within LangChain middleware to selectively gate specific dangerous tools. This allows granular control over which tool calls require human review, editing, or rejection before execution.

Structured output validation in LangChain agent workflows is handled using Pydantic or Zod schemas. This ensures that tool call outputs and resumed commands adhere to defined data structures before continuing the pipeline.

Persistent context in LangChain human-in-the-loop pipelines is maintained using thread-scoped state with checkpointer support. This ensures that workflow contexts remain intact across paused tool calls and subsequent resume commands.