Researchers at the University of Florida have built a lightweight system designed to keep autonomous AI agents focused on their assigned jobs. The system, called Algebraic Task-role Locking with Adaptive Spectral Memory, or ATLAS, was developed by computer science professor Sumit Jha and his team. They plan to present the research at the 2026 Conference on Empirical Methods in Natural Language Processing in Budapest.
Autonomous AI systems are taking on more complex tasks independently, but keeping them focused remains a major challenge. Recent incidents, including AI models accessing restricted government websites and sending odd anonymous police tips, point to the risks of unpredictable agent behavior.
How ATLAS differs from existing AI safety checks
Most safety systems today review an AI’s answers only after they have been generated. ATLAS works differently, operating inside the neural network while the model produces its output. Jha said a model can begin drifting from its assigned role before it gives a wrong answer. The system detects this confusion early and guides the agent back, without slowing things down or requiring large amounts of extra computing power.
Role decoherence in long-running AI agent teams
This internal breakdown, which the team calls role decoherence, often occurs when groups of AI agents work together over long periods. As an example, a model acting as a medical professional might gradually lose sight of its main job after handling millions of data points over several days. ATLAS uses mathematics to reinforce the core mission, the agent’s specific duties and the current facts, helping it stay focused during extended tasks.
Jha compared the approach to ongoing medical scans of a neural network in real time. If the system sees the agent drifting from its assigned medical tasks, the software calculates the exact changes needed to correct it. Because these adjustments are small, the AI can be returned to its task before a minor distraction becomes a serious problem.
