Autonomy is a loop, not a model choice
Integrated sensing and communications already connects radio observations with communication decisions. A new survey on agentic ISAC argues that adding an agent changes the unit of analysis: the system must be evaluated from observation through reasoning and action, then back through feedback and resilience. The authors call this broader paradigm AISAC.
Their framework divides the loop into six stages: observation; contextualization; reasoning and prediction; planning and orchestration; execution and collaboration; and feedback and resilience. This structure places physical sensing, multimodal interpretation, reinforcement learning, large language models, reconfigurable surfaces, edge intelligence, and multi-agent coordination inside one operational cycle. Security, privacy, resilience, and sustainability are treated as cross-cutting requirements rather than features that can be attached after the controller is built.
Evidence must scale with the autonomy claim
The survey also proposes five maturity levels, beginning with physical-layer primitives and ending with fully closed-loop agentic ISAC. A maturity scale is useful because a learned detector, an optimizer, and an autonomous multi-agent system should not receive the same label. The required evidence grows with the claim: a component can be tested for task accuracy, while a closed-loop agent must also be tested for tool safety, timing, recovery, coordination, and behavior under distribution shift.
The sharpest result in the public abstract is an audit of representative studies against nine agent-specific evaluation criteria. No reviewed system reports more than one or two. That observation does not mean the underlying sensing or communication methods are ineffective. It means the evidence normally reported for a component is not enough to establish mature autonomy for the complete loop.
The practical value of the survey is therefore its evaluation discipline. Agentic wireless systems need benchmarks that expose where context comes from, which actions are permitted, how real-time physical-layer constraints enter planning, and what happens when an action fails. Without those measurements, “agentic” remains an architectural aspiration. With them, it becomes a falsifiable systems claim.
Research notes
When Agentic AI Meets Integrated Sensing and Communication
- Authors: Kai Li, Conggai Li, Sarah Ali Siddiqui, Syed Sohail Ahmed, Xin Yuan, Shenghong Li, Wei Ni
- Public record: arXiv
- What is established: The survey defines a six-stage closed loop, five maturity levels, and nine agent-specific evaluation criteria spanning sensing, reasoning, action, collaboration, and resilience.
- Read with care: The audit is a taxonomy-driven evidence review; it does not itself demonstrate a deployed fully autonomous ISAC system.