ARIA Enables Runtime Repair for Multi-Agent AI Systems

Summary

In large language model-based multi-agent systems, an early local error can spread as later agents rely on it, causing an otherwise recoverable task to fail. The paper introduces Agent Runtime Inspection and Amendment (ARIA), a framework for learning when to intervene during execution and how to continue after a repair. The authors collect more than 24,000 successful and failed trajectories across four multi-agent frameworks, six benchmarks, and eight host models. They use these trajectories to build what they describe as the first large-scale dataset pairing failed multi-agent runs with execution-verified repairs. The dataset contains 3,953 repair pairs that identify the failure, the intervention point, and the required continuation. ARIA is designed to work with any multi-agent system that exposes execution trajectories and supports replacing the continuation of a run. A compact model is trained on the verified repairs to recognize intervention opportunities and produce corrections that can transfer to related failures, rather than memorizing individual fixes. The trained system shows strong repair performance in both post-hoc and runtime settings, including generalization to frameworks and host models not seen during training. The authors provide the implementation publicly on GitHub.