CORRECT: COndensed eRror RECognition via knowledge Transfer in multi-agent systems

Multi-agent systems (MAS) are increasingly capable of tackling complex real-world tasks, yet their reliance on inter-agent coordination, tool use, and long-horizon reasoning makes error recognition particularly challenging. Minor errors can propagate across agents, escalating into task failures while producing long, intertwined execution trajectories that impose significant costs for both human developers and automated systems to debug and analyze. Our key insight is that, despite surface differences in failure trajectories (e.g., logs), MAS errors often recur with similar structural patterns. This paper presents CORRECT, the first lightweight, training-free framework that leverages an online cache of distilled error schemata to recognize and transfer knowledge of failure structures across new requests. This cache-based reuse allows LLMs to perform targeted error localization at inference time, avoiding the need for expensive retraining while adapting to dynamic MAS deployments in subseconds. To support rigorous study in this domain, we also introduce CORRECT-Error, a large-scale dataset of over 2,000 annotated trajectories collected through a novel error-injection pipeline guided by real-world distributions, and further validated through human evaluation to ensure alignment with natural failure patterns. Experiments across seven diverse MAS applications show that CORRECT improves step-level error localization up to 19.8% over existing advances while at near-zero overhead, substantially narrowing the gap between automated and human-level error recognition.

Large language models are increasingly being used in multi-agent systems, where several AI agents work together to solve complex tasks such as software development, web navigation, or scientific reasoning. However, when these systems fail, it is often difficult to know which agent made the first important mistake, because one small error can spread through later interactions and lead the whole system to fail. This makes debugging multi-agent systems slow, expensive, and hard to scale. In this work, we introduce CORRECT, a lightweight method for recognizing the decisive error in a failed multi-agent run. Instead of retraining a model or asking it to inspect many long past failure logs, CORRECT summarizes previous failures into compact error patterns, stores them, and retrieves the most relevant ones when a new failure occurs. These patterns help a language model focus on the most likely source of the problem. We also build CORRECT-Error, a large benchmark of annotated multi-agent failures. Across diverse tasks, CORRECT improves error localization while adding little overhead, helping make multi-agent AI systems easier to debug, monitor, and deploy reliably