In large‑language‑model (LLM) based multi‑agent systems (MAS), a local mistake made by one agent can be propagated by others, turning a correctable error into a task failure. Conventional approaches either diagnose after execution or repeatedly re‑run the whole system to locate fixes, which is time‑consuming and does not yield a learnable policy for intervening during runtime.
The Agent Runtime Inspection and Amendment (ARIA) framework addresses this by converting execution‑verified repairs into reusable supervision. Repair search is shifted to data construction and model training, allowing a trained model to intervene on‑the‑fly without exhaustive testing of alternatives.
We first gathered more than 24,000 successful and failed trajectories across four MAS frameworks, six benchmarks, and eight host models. From these we built the ARIA dataset—the first large‑scale collection that pairs failed multi‑agent trajectories with execution‑verified repairs. The dataset contains 3,953 verified repair pairs, each specifying the error type, the intervention point, and the correct continuation of execution.
Using this dataset, ARIA provides a MAS‑agnostic runtime inspection and amendment interface: any MAS that can emit execution trajectories and support replacement of the remaining execution can be integrated. A compact model is trained on the verified repairs to recognize when intervention is needed and how to amend an ongoing run. The goal is to learn general repair strategies that transfer to similar failures in new configurations, rather than memorizing individual fixes.
Experiments demonstrate that the trained ARIA model achieves strong repair performance in both post‑hoc and runtime settings, and it generalizes well to unseen frameworks and host models. The code is released at https://github.com/LabRAI/ARIA.
Review: ARIA pushes repair search to the data level and leverages large‑scale verified fixes to train a universal corrective model, enabling real‑time self‑healing for LLM‑driven MAS and offering a practical path toward more robust multi‑agent deployments.