rubricmiddleware This name sounds like it was made by the academic affairs office, but it really hits the pain point: the agent’s earlier halfhearted attitude of “I think it’s done” can now be hard-controlled by rules until it’s truly finished.

CoinWorld News reports that LangChain has released a new deep agents component, rubricmiddleware, aiming to enable AI agents to check and revise their own outputs according to predefined standards. Developers can set task “completion criteria,” such as requiring the code to pass tests, the report to cover specified sections, and the answer to not include prohibited content. Each time an agent is preparing to deliver a result, the system calls a review model to check items one by one; if the criteria are not met, the feedback is sent back to the original agent to continue making modifications until it passes the checks or reaches the iteration limit. This mechanism addresses the common issue of “missing the last step” when agents carry out long tasks, ensuring the agent knows what truly counts as completion.