ARIA: Runtime Inspection and Amendment for Multi-Agent Systems

ARIA is a new framework designed for Runtime Inspection and Amendment in LLM-based multi-agent systems. It focuses on learning how to repair ongoing executions to prevent local mistakes from propagating into task failures. Meanwhile, splitting a single agent into several provides specific benefits, primarily separate context windows and parallelism. This approach prevents context overflow by allowing subagents to handle specific research, though it can increase costs and complicate debugging if not applied to broad tasks.