As artificial intelligence systems become increasingly capable of working on their own, keeping them focused on their designated tasks is a growing challenge.
University of Florida computer science professor Sumit Jha, Ph.D. , is working to keep agents from "going rogue." Jha and his team have developed a system that can detect when an AI model begins to drift from its assigned task and steer it back without retraining the model or adding significant computing demands.
The research comes as news reports have chronicled AI agents gaining more autonomy and occasionally behaving in ways their developers did not intend. Recent incidents have included AI agents accessing government websites and an Anthropic model submitting an anonymous tip to a police department in July, claiming to have information about an unsolved murder.
Jha, a professor in UF's Department of Computer & Information Science & Engineering, or CISE , studies how to make AI systems more controllable and understandable. He is working to engineer simple, non-invasive tweaks to keep systems on task.
Jha and his collaborators have authored a paper entitled "Drift No More: Spectral Locking for Long-Horizon Open-Weight Multi-agent AI," detailing a system, dubbed ATLAS, which stands for Algebraic Task–role Locking with Adaptive Spectral Memory.
The acronym hints at the dense mathematics at the heart of the system.
Jha and his co-authors, Vishal Pramanik, Maisha Maliha and Susmit Jha, devised a system that works inside the large language model, or LLM, as it self-generates. When the model begins to drift from its task, the ATLAS system steers the LLM back toward the assignment, as well as its original responsibility and the latest shared information.
"An AI agent can start going wrong before its response looks wrong," Jha explained. "Inside the model, it may already be drifting from the task, confusing its role or relying on information that has been replaced. ATLAS is built to correct that drift before it reaches the next message, so other agents do not build on the mistake."
Existing strategies usually involve policing an AI agent's answers after the fact. The ATLAS system doesn't require retraining the model, and it therefore doesn't expend extra computing resources or slow the model down.
According to Jha, so-called task drift is especially possible for tasks in which groups of AI agents coordinate with each other over long periods of time.
"Say, initially you told the agent you're a doctor, you are supposed to treat patients," Jha said. "By day No. 7, the fact that it's a doctor and it should treat patients is now just one of the millions of things it has been told. We call this problem role decoherence. The agent was tasked to do something, but after some time it has been told so many other things, it loses track."
AI teams are usually managed from the outside through prompts, retrieval systems and rules, Jha said.
"ATLAS adds something closer to an internal compass," he said. "In mathematical form, it continually reinforces three questions: What is the mission? What is this agent responsible for? And what is true right now? Keeping those three answers aligned is essential when a task lasts much longer than a single conversation."
Sounds complicated, but as Jha tells it, the system is akin to taking MRI snapshots of the AI's brain, or neural network. Imagine an AI model that is supposed to be performing the tasks of a medical professional.
"Initially when I do the MRI, it looks like the brain of a doctor," he explained. "Thumbs up. But when we do another MRI, now it looks like, let's say, the brain of a race car driver. So, we look inside and we find out what the difference was between the old MRI and the new MRI."
As ATLAS had already learned what the medical neural network should look like, it is able to calculate the smallest set of changes necessary, within the neural network, to return the system to its original task of medical assessment.
"The agent has already launched, so the intervention has to be very small," Jha explained. "Taking the MRI is like taking a snapshot, it doesn't require retraining, so you don't have to stop the agent. If the agent's doing alright, it's still a doctor. ATLAS doesn't do anything. But if it's now 99% a doctor, but 1% race car driver, we take the 1% out. We don't wait until it's a problem."
The paper has been accepted to the 2026 Conference on Empirical Methods in Natural Language Processing, or EMNLP , Oct. 24–29 in Budapest.
