LangChain adds an auto quality inspector to the Agent. If the task cannot be completed, it will be returned for revision.

According to Pulse Check Beating monitoring, LangChain has released a new Deep Agents component called RubricMiddleware, which allows AI Agents to inspect and modify their own output against predefined criteria. Developers can clearly define the "completion criteria" for a task, such as ensuring code passes tests, reports cover specific sections, and responses do not contain prohibited content. Each time an Agent is ready to deliver results, the system invokes a review model to check each criterion; if any criterion is not met, feedback is sent back to the original Agent for further modification until it passes the checks or reaches the iteration limit.

This mechanism addresses the common issue faced by Agents undertaking long tasks of "falling short at the final step." Many Agents are not incapable of completing the task but often overlook hard requirements such as formatting, testing, references, and sections. The RubricMiddleware serves as an automated quality control officer in the task workflow, guiding the Agent on what constitutes true completion rather than just producing a seemingly acceptable answer.

The LangChain documentation explicitly states that this approach is most suitable for tasks with clear acceptance criteria, such as ensuring the correct syllable count in a haiku, passing tests after code refactoring, or including all necessary sections in a report. For the average user, the value lies not in making the Agent better at chatting but in making the Agent more like an executor who can deliver results based on a checklist.