UF researchers develop ATLAS to steer drifting AI agents back on task

Keeping an AI agent on task may require intervening before its answer looks wrong. University of Florida professor Sumit Jha and collaborators have developed ATLAS to do that inside a language model as it generates a response. Their claim is ambitious: correct the agent’s drift without retraining it or significantly increasing computing demands.

The university described the work in an October 29 research announcement. The 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, or EMNLP. The conference runs October 24–29 in Budapest.

How an assignment gets lost

Jha says task drift is especially relevant when groups of AI agents coordinate over long periods. The problem is not just an incorrect answer. An agent may lose track of its responsibility, wander away from the assignment or keep relying on information that newer shared information has replaced.

He illustrates this with a hypothetical agent instructed to act as a doctor and treat patients. After days of additional instructions and exchanges, that original assignment becomes one item among a vast amount of information. The agent can then lose its sense of the role—a problem Jha calls “role decoherence.”

That matters for a team because one agent’s next message can become material another agent uses. ATLAS is designed to intervene before that handoff, rather than let other agents build on a mistake. Its target is the developing response, not just the answer that eventually becomes visible.

An AI agent can start going wrong before its response looks wrong. ATLAS is built to correct that drift before it reaches the next message, so other agents do not build on the mistake.

Sumit Jha, University of Florida computer science professor

A small correction inside the model

ATLAS stands for Algebraic Task–role Locking with Adaptive Spectral Memory. Jha describes it as an internal compass, contrasting it with managing agents from outside through prompts, information-retrieval systems and rules. The system is intended to keep three things aligned during generation:

• The mission: the task the agent is supposed to complete.
• The role: the responsibility assigned to that particular agent.
• The current information: what the team should treat as true now.

To explain the mechanism, Jha compares examining the neural network to taking MRI snapshots. In his medical-agent analogy, an initial snapshot reflects the assigned role. A later snapshot reveals whether the model has moved away from it. ATLAS then calculates the smallest changes needed to steer it back.

The intervention is meant to be selective. Jha says ATLAS does nothing when the agent remains on course; it makes a small correction when drift appears. Because the agent is already running, he says, the adjustment must not require stopping it for retraining. The researchers also say the system does not slow the model.

The design claim is clearer than the measured result

The university’s account supplies no quantitative evaluation of ATLAS: no measured reduction in drift and no timing or computing-overhead figures. Its effectiveness and performance therefore remain researcher claims here, rather than results readers can assess numerically. Conference acceptance does not fill that evidentiary gap.

Jha developed the work with co-authors Vishal Pramanik, Maisha Maliha and Susmit Jha. Their paper addresses long-running, open-weight multi-agent systems. The announced contribution is a way to detect and correct drift during generation—not a demonstrated guarantee that autonomous agents will always stay within their assignments.