ATLAS steers drifting AI agents back on task from inside the model, University of Florida team says
The system needs no retraining of the model, and its creators say it adds no significant computing load, though no measured figures were published in the release.
In brief
A fix for AI agents that wander off task
University of Florida's Sumit Jha built ATLAS. It spots drift inside the model and steers it back. No retraining is needed. The release gives no measured results yet.
A system called ATLAS can tell when an AI model starts to wander from its assigned job and nudge it back. It was built by University of Florida computer science professor Sumit Jha, Ph.D., and his team. We know about it from a UF release and a paper in the EMNLP 2026 proceedings.
What it does
An AI agent can start going wrong before its response looks wrong.
ATLAS works inside the language model while it writes its output. When it senses drift, it pulls the model back toward three things: the task, the agent's original role, and the latest shared information.
The team says this needs no retraining of the model and no significant extra computing.
Why it matters
AI agents are programs that act on their own toward a goal. The more freedom they get, the more it matters that they stay on course.
The UF release points to news reports of agents behaving in ways developers did not intend. It cites agents accessing government websites, and an Anthropic model sending an anonymous police tip in July about an unsolved murder.
Imagine a team of agents planning a trip together. One forgets it is the budget checker and starts booking hotels. The others then build on that mistake.
That is the risk Jha describes. In his words, an agent can start going wrong before its response looks wrong. ATLAS aims to fix the drift before it reaches the next message.
Background
The paper is titled "Drift No More: Spectral Locking for Long-Horizon Open-Weight Multi-Agent AI." The authors are Vishal Pramanik, Maisha Maliha, Susmit Jha and Sumit Kumar Jha. ATLAS stands for Algebraic Task-role Locking with Adaptive Spectral Memory.
The release says most existing approaches check an agent's answers after the fact. ATLAS acts earlier, during generation. One outlet, World Today Journal, calls the underlying problem "role decoherence."
According to Jha, drift is most likely when groups of agents coordinate over long periods.
Limits
The university says ATLAS does not slow the model and makes only small corrections. The university account gives no measured figures for drift reduction, timing or computing overhead. The results inside the paper were not available in the material reviewed. Independent coverage is thin, and no source describes a public release or next steps.