ATLAS aims to keep autonomous agents on task

University of Florida computer science professor Sumit Jha and his collaborators have developed ATLAS, a system designed to detect when an AI agent strays from its assignment and redirect it before the mistake appears in its next message. The approach works inside a large language model as it generates responses, rather than policing its answers afterward.

ATLAS reinforces the agent’s mission, assigned responsibility and latest shared information. Jha says this alignment is especially important when groups of agents coordinate over long periods, accumulating instructions and updates that can cause them to lose track of their original roles or rely on outdated information. Correcting drift internally is intended to prevent other agents from building on an emerging mistake.

The system compares the model’s internal state with its intended task and calculates small changes needed to restore alignment. It leaves the model alone when it remains on track. According to Jha, the intervention requires no retraining, does not slow the model and avoids significant additional computing demands. The team’s paper, “Drift No More: Spectral Locking for Long-Horizon Open-Weight Multi-agent AI,” has been accepted to the 2026 Conference on Empirical Methods in Natural Language Processing.