LangChain adds an automatic quality inspector to the agent, and if it cannot be completed, it will be returned for rework According to Beating monitoring, LangChain has released a new component called RubricMiddleware for Deep Agents, which allows AI agents to check and modify their output according to preset standards. Developers can first write clear 'completion criteria' for tasks, such as code passing tests, reports covering specified chapters, and responses not containing prohibited content. Every time the agent prepares to deliver results, the system will call a review model to check each item one by one; As long as the standard is not met, feedback will be returned to the original agent for further modification until it passes the inspection or reaches the iteration limit. This mechanism solves the common problem of "missing the last step" when agents perform long tasks. Many agents are not completely incapable of doing it, but are prone to missing hard requirements such as formatting, testing, referencing, and chapters. RubricMiddleware is like adding an automatic quality inspector to the task chain, allowing the agent to know what truly counts as completed, rather than just generating a seemingly similar answer. The LangChain documentation also clearly states that this approach is most suitable for tasks with clear acceptance criteria, such as whether haiku syllables are correct, whether code refactoring tests pass, and whether the report includes all necessary parts. For ordinary users, its value lies not in making the agent chat better, but in making the agent more like an executor who can perform tasks according to a checklist. [Original link]