This paper has been accepted for publication at IEEE MILCOM Workshops 2026. This is the author’s accepted version of the paper.
Agentic RDZ: Autonomous Zone Management with AI Agents and an FR3 Coexistence Use Case Minh Dat Nguyen, Gabriele Gemmi, Tamerlan Aghayev, Paolo Testolina, Michele Polese, Tommaso Melodia
arXiv:2609.17110v1 [cs.NI] 15 Sep 2026
Institute for Intelligent Networked Systems, Northeastern University, Boston, MA, U.S.A. {minhd.nguyen, g.gemmi, aghayev.t, p.testolina, m.polese, melodia}@northeastern.edu
Abstract—Radio Dynamic Zones (RDZs) allow wireless experiments to operate outside conventional spectrum regulations while continuously guaranteeing protection for incumbent users. Existing RDZ prototypes automate this task procedurally, through handcrafted rules and predefined workflows, and become brittle when experiments encounter hardware impairments, user workflows and devices, or interference mechanisms not anticipated at design time. This paper introduces the agentic RDZ (A-RDZ), which, to the best of our knowledge, is the first RDZ realization in which agents use Large Language Models (LLMs) to perform spectrum management, experiment management, policy interpretation, and zone orchestration. Built on the G ENESIS agentic framework, the architecture pairs autonomous reasoning with a deterministic policy gate and near-real-time (near-RT) reflexes, so that agents can improve outcomes but never weaken the zone’s protection guarantee. We validate the A-RDZ on a hardware-inthe-loop Frequency Range 3 (FR3) (7.125–24.25 GHz) Open Radio Access Network (O-RAN) testbed in which a 5G New Radio (NR) experiment coexists with an emulated Fixed Satellite Service (FSS) earth-station incumbent. In an end-to-end use case, the monitoring agent detects an emission violation from live spectrum evidence, the orchestrator selects a mitigation that restores the interference budget while keeping the experiment running, and the action is applied and verified through the O-RAN control plane. We report the detection-to-mitigation latency decomposition and discuss the practical limits of agentic operation, including non-deterministic reasoning and decision-to-action translation. Index Terms—Radio Dynamic Zones, agentic AI, LLM agents, O-RAN, FR3, spectrum coexistence, testbeds.
I. I NTRODUCTION Future 6G systems are expected to exploit the Frequency Range 3 (FR3) upper mid-band (7.125–24.25 GHz) for wider bandwidths than conventional sub-7 GHz systems [1], [2]. However, FR3 is already occupied by incumbents such as Fixed Satellite Service (FSS) earth stations and federal systems, making coexistence a key deployment challenge. Characterizing that coexistence requires over-the-air (OTA) measurements with real hardware and protocol stacks to account for non-ideal radio frequency (RF) behaviors and propagation. Experimental authorization is granted against pre-declared parameters at a fixed site. Highly-dynamic or wide-ranging tests, however, have requirements that are hard to predict, or change based on the outcome of the experiment itself. In that case, obtaining a new license might significantly slow down the experiment pace and the scientific output. This work was supported by the U.S. National Science Foundation under Grants CNS-2434081 and OSI-2431961.
Radio Dynamic Zones (RDZs) address this limitation by providing controlled environments for spectrum experimentation, enabling wireless systems to operate under dedicated experimental authorizations while continuously protecting incumbent users [3]. Existing systems, including OpenZMS on POWDER [4], [5], COSMOS [6], and FlexRDZ [7], automate spectrum, experiment, and policy management through predefined rules, workflows, or task models. Their coverage is therefore bounded by what the designer explicitly specified: a hardware impairment or interference mechanism outside that enumeration has no matching rule, and the zone’s remaining response is to terminate the experiment. Recent agentic artificial intelligence (AI) systems provide complementary capabilities: they can reason over heterogeneous evidence and invoke tools [8], recover from failures through self-correction [9], and coordinate multi-step operations [10]. G ENESIS [11], for example, demonstrated multiagent autonomous operation of Open Radio Access Network (O-RAN) testbeds. These capabilities motivate applying agentic AI to RDZ operation, while retaining deterministic mechanisms for safety-critical protection. Motivated by these observations, we propose the agentic RDZ (A-RDZ), which, to the best of our knowledge, is the first extension of the conventional RDZ architecture with coordinated AI agents for spectrum, experiment, and policy management and zone orchestration. The architecture supports AIbased decisions, guard-railed by deterministic policies and realtime protection mechanisms. We validate it on a hardware-inthe-loop FR3 O-RAN testbed, where a 5G New Radio (NR) experiment coexists with an emulated FSS incumbent. During the experiment, a controlled configuration change pushes the unwanted emissions into the protected sub-band. The monitoring agent detects that the measured interference exceeds the protection limit and identifies the inverting frequency plan as the cause, rather than the transmitter’s in-band configuration. The orchestrator then selects a Physical Resource Block (PRB)region mask that brings the interference back within the limit, while the experiment continues to run. The main contributions are: We introduce the concept of an agentic RDZ (A-RDZ), extending conventional RDZ functions with coordinated AI agents while retaining deterministic safety mechanisms. • We integrate A-RDZ with the G ENESIS framework and a hardware-in-the-loop FR3 O-RAN coexistence testbed, •
TABLE I RDZ OPERATIONAL CHALLENGES MAPPED TO AGENTIC CAPABILITIES AND TO THE DETERMINISTIC GUARDS THAT BOUND AGENT BEHAVIOR . agents
Deterministic safeguard
Agentic capability
Spectrum management
Reflex threshold test and pre-configured protective action
Evidence-based anomaly reasoning and predictive violation detection
guard
Experiment management
Grant validation before execution
Autonomous deployment, diagnosis, and recovery
management
Policy management
Non-overridable protection limits
Reasoning over structured and natural-language policies
Decision engine
near-real-time (near-RT) reflexes (10–1000 ms); agents (1–200 s)
Cross-subsystem coordination and mitigation planning
II. RDZ O PERATIONAL C HALLENGES AND AGENTIC S UPPORT An RDZ consists of three functional subsystems coordinated by a decision engine [3]: Spectrum Management, Experiment Management, and Policy Management. We summarize the key operational challenges of each subsystem and discuss how an agentic realization complements conventional procedural automation. Table I summarizes the mapping. • Spectrum Management: Spectrum management must detect interference violations and anticipate harm to protected incumbents. Conventional systems rely on threshold checks derived from propagation and emission models, which may fail when experimental hardware behaves unexpectedly. An agent can instead correlate spectrum measurements with experiment configurations and historical observations to attribute anomalous emissions and predict violations before protection limits are exceeded. It can also assess whether a measured exceedance causes harmful interference at the incumbent. Each individual computation is scriptable and is already implemented in existing systems. The agent is useful when the script has no matching case: it can assess an incumbent scenario, geometry, or interference path not covered by the encoded rule set, and defers to the deterministic threshold when no such assessment can be made. • Experiment Management: Experiments execute on shared radio and compute infrastructure, requiring reliable deployment, monitoring, recovery, and cleanup. While scripts automate nominal workflows, diagnosing failed deployments often requires reasoning over logs, system state, and configurations. Moreover, agentic management supports experiment workflows outside the encoded set, including mapping a new workflow onto the spectrum-management requirements it implies. Deterministic policy checks keep all lifecycle actions within the experiment’s authorized spectrum grant.
occupants DB · blueprints episode traces
(decision engine)
RDZ function
demonstrating agentic detection, mitigation selection, and policy-gated execution. • We evaluate the end-to-end enforcement loop on several metrics, including latency, protection compliance, experiment utility, and the role of deterministic safeguards under agentic operation.
Zone knowledge base
Zone Orchestrator
Spectrum Sentinel
Experiment Steward
Policy Interpreter
actuate grant Policy Gate (G ENESIS hooks)
max hard limits: IFSS , EIRP, grant, boundary
verified action Management & Control RIC/SMO · OAM · vendor API
control · lifecycle substrate
Baseband L1–L3
provision mask
enforce mask
Reflex bounded-time backstop
RF control Radio unit
RF front-end
Protected incumbent
sensing evidence
+ sensing node
Fig. 1. The agentic RDZ architecture: G ENESIS agents (top) realize the RDZ functional components of [3].
Policy Management: Zone policies combine structured spectrum rules with natural-language regulatory documents and coordination agreements. Procedural systems require these policies to be manually encoded into machine-parsable text or code before enforcement. A Large Language Model (LLM)-based agent can instead interpret heterogeneous policy documents, directly evaluate natural-language ones, evaluate experiment requests, and recommend the leastdisruptive compliant mitigation. Hard protection constraints remain enforced by a deterministic policy gate. • Decision Engine: The decision engine coordinates the three subsystems while satisfying stringent protection timescales. Since LLM inference cannot guarantee sub-second response, the A-RDZ adopts a two-timescale architecture: deterministic near-RT reflexes provide immediate protection, while agents operate at slower timescales to diagnose anomalies, select mitigations, and coordinate system-wide actions. Agents can thus increase the autonomy and automation of RDZs by interpreting heterogeneous information and recovering from unforeseen operating conditions. However, agentic operation introduces two key challenges: non-deterministic reasoning and imperfect decision-to-action translation. Accordingly, safety guarantees are required. Deterministic policy gates and near-RT reflexes, enforced through typed interfaces and action verification routines, limit the agent errors. •
III. AGENTIC RDZ A RCHITECTURE Figure 1 shows the A-RDZ architecture. We implement it by building on G ENESIS [11], an agentic framework for ORAN testbeds that provides domain-scoped agents, reusable skills, and deterministic hooks that can block unsafe actions before network execution, to implement and extend the RDZ functional components of [3] The architecture is organized into four planes. An agent plane performs the reasoning functions of the zone; a guard plane enforces the zone’s guarantees through deterministic protections; a management plane translates verified decisions into network configuration; and a substrate plane includes the managed radio system, the protected incumbent, and the sensing nodes that observe it. The management plane provides two control paths: an agentic path that applies only decisions
verified by the guard plane, and a deterministic reflex that applies a pre-configured protective action when the protection threshold is crossed, independently of agent liveness. The management plane is the sole component permitted to write to the substrate, making the architecture independent of the interface used to reach it: O-RAN E2/O1, 3GPP Operations, Administration, and Maintenance (OAM), or a device-specific application programming interface (API). The zone knowledge base stores the occupants database, the measurement and policy blueprints, and the enforcement history. It may be specific to a zone, or federated with an experimenter’s knowledge base. Specialist Agents: Following the G ENESIS design, in which each agent is scoped to a single domain and given only the tools and authority required for that domain, three specialist agents realize the core RDZ subsystems. – Spectrum Sentinel (spectrum management): The Spectrum Sentinel manages the sensing nodes and the measurement blueprint. It implements a traditional interference threshold test as a reflex: a deterministic check that requires no reasoning. The Sentinel agent extends it by testing whether the observed spectrum matches the declared configuration. From the carrier, occupied bandwidth, PRB allocation, and frequency plan, it determines where emitted power should appear. A change in the occupied region indicates a configuration change; energy in the adjacent sub-band predicted by the injection side indicates a frequency-plan artifact; and energy where no resources are allocated indicates that the transmitter is not operating as declared. It also assesses whether a limit breach can cause actual harm at the incumbent, and flags emissions inconsistent with the declared configuration while still within budget. Its outputs are detected and predicted violations. – Experiment Steward (experiment management): The Experiment Steward manages the occupants database and experiment lifecycle. It instantiates an experiment from its spectrum grant and declared configuration, then verifies the resulting emissions against the grant. During the run, it applies reconfigurations, recovers from common failures, and executes orchestrator-selected mitigation actions. It tears the experiment down at completion or following an enforcement decision. All lifecycle operations are validated by the policy gate before execution. – Policy Interpreter (policy management): The Policy Interpreter manages the zone policy blueprint, including protection criteria, coordination agreements, and admission rules. It translates experiment requests into enforceable spectrum grants, evaluates mitigation actions for policy compliance, and maintains an auditable record of policy decisions. • Zone Orchestrator (decision engine): The Zone Orchestrator implements the RDZ decision engine. It is a deterministic component rather than an agent: it has no LLM of its own and invokes the specialist agents as typed functions. On a violation, it collects the attribution, candidate mitigations with projected post-action interference and disruption cost, and •
the protection target, then selects the lowest-cost candidate that meets the target. The selected action is delegated through the policy gate and verified by an independent measurement. The agents provide attribution and candidate mitigations; the orchestrator makes the final selection. At admission, it also derives the protection threshold and pre-computed reflex action from the grant. All decisions and outcomes are recorded in the knowledge base. • Guard Plane: All actuation originating in the agent plane passes through a policy gate: a deterministic validator holding the zone’s hard limits, the incumbent interference budget max IFSS derived from the applicable protection criterion, absolute Effective Isotropic Radiated Power (EIRP) caps, the frequency grant of each experiment, and the geographic boundary. The gate rejects any action that would result in a violation, and only actions it has verified are admitted to the management plane. It follows the reference-monitor model [12] and, for learned policies, the shielding pattern [13], in which a correct-by-construction monitor overrides any agent action violating a safety specification. In the G ENESIS substrate, this is realized as hooks that fire out-of-band of the agents’ reasoning loop and can hard-block an action before it reaches the network [11]. The gate deterministically evaluates the expected outcome from the action parameters and the current calibrated measurement: a PRB mask removes the measured contribution of the selected PRBs, while a power offset scales the emission by a fixed factor. The agent only proposes the action; the gate independently computes and checks its expected outcome. • Management Plane: The management plane converts a verified action into concrete configuration on the substrate and arbitrates between the two timescales at which the zone acts. A verified PRB-mask action, for example, becomes an ORAN RAN Intelligent Controller (RIC) control message in our testbed, a scheduler configuration change under 3GPP OAM, or a front-end gain reduction where no networkcontrol path exists. It is the sole point at which the architecture writes to the network and therefore the only place specialized to a particular technology. The management plane also hosts the reflex: a pre-configured, bounded-time protection path that evaluates the monitored interference budget against the protection threshold established at experiment admission and applies a pre-computed protective action, such as a PRB mask, when the monitored budget is exceeded. It operates independently of agent liveness and performs no LLM inference, so its response time is bounded by the sensing cadence and the control interval of the underlying network rather than by reasoning latency. • Substrate Plane and Sensing: The substrate plane comprises the managed radio, processing stack, RF front-end, protected incumbent, and sensing nodes. The management plane exposes network-control actions (e.g., PRB masking, power adjustment, and Modulation and Coding Scheme (MCS) capping), lifecycle-control actions (container, process, and configuration management), and RF-control ac-
A-RDZ management plane Zone Orchestrator · Spectrum Sentinel · Experiment Steward · Policy Interpreter E2 KPM / ctrl lifecycle tools
gNB host OAI nr-softmodem, 7.2x near-RT RIC + reflex xApp N2/N3
policy gate
FR3 OTA · Vivaldi Tx/Rx
Foxconn O-RU
IF chain att. / filt. / tap
IF
Pi-Radio
Pi-Radio
IF → FR3
FR3 → IF
IF
IF chain
UE: USRP B210
att. / filt. / tap
OAI nr-uesoftmodem
virtual FSS victim
calibrated PSD evidence
IF tap
xRAN FH VLAN C/U
FieldFox N9953B
Open5GS core
IF tap
calibrated I estimate
S.1428 pattern · S.1432 I/N
calibrated monitor
Fig. 2. FR3 O-RAN testbed and its exposure to the A-RDZ management plane.
tions such as transmit-gain adjustment. Sensing is a northbound observation path: calibrated power spectral density (PSD) captures reach the Spectrum Sentinel without traversing the management plane. Zone-boundary sensing is left to future work in this single-link realization. IV. E XPERIMENTAL S ETUP AND G ENESIS /FR3 I NTEGRATION A. FR3 O-RAN Testbed The experiment managed by the A-RDZ (Fig. 2) is a standalone 5G NR network built on OpenAirInterface (OAI) [14]. The next-generation NodeB (gNB) runs nr-softmodem with a 7.2x functional split in a Docker container, driving a Foxconn O-RAN Radio Unit (O-RU) over xRAN fronthaul via an Intel X710 network interface card (NIC). The cell operates in band n78 with 106 PRBs at 30 kHz subcarrier spacing (SCS) (µ = 1) and fIF = 3.58 GHz. The Frequency Range 1 (FR1)/intermediate frequency (IF) path uses calibrated attenuators, filters, and splitters with monitoring taps for repeatable coupling. A pair of Pi-Radio up/down-conversion boards [15] translates the IF to FR3 using a common fLO = 13.58 GHz with high-side injection, inverting the spectrum and placing the dominant emission skirt in the lower-adjacent sub-band. The FR3 segment forms a short indoor OTA link between directive Vivaldi antennas, with the fixed geometry included in the endto-end calibration. A USRP B210 running nr-uesoftmodem serves as the user equipment (UE) behind the down-converter, while Open5GS provides the core network on a separate host.
and delivered to the Spectrum Sentinel; their capture cadence contributes to the enforcement-loop latency. C. Agent Integration The A-RDZ management plane runs off-testbed and reaches the infrastructure through three tool families, each exposed to the agents as a typed, schema-validated function and protected by the policy gate: • RIC tools: Key Performance Measurement (KPM) subscription for per-UE/per-slice telemetry and Radio Access Network (RAN)-control actions (PRB-region mask, power offset, and MCS cap) issued via an xApp on the near-RT RIC using FlexRIC [18]. • Lifecycle tools: container control (deploy from the committed gNB image, stop, commit, log retrieval), configuration templating for the OAI and O-RU config files, and host hygiene actions (fronthaul virtual function (VF) re-initialization with the pinned MAC address, stale hugepage-map cleanup) that the Steward runs before any deployment. • Sensing tools: triggered and periodic capture from the FieldFox, which also realizes the virtual victim observation point, returning calibrated PSDs and scalar budget estimates. The reflex is an xApp configured at experiment admission: it evaluates the interference-budget estimate at each control interval (∆t = 100 ms) and applies a pre-computed protective PRB max , independently mask if the instantaneous estimate exceeds IFSS of agent liveness.
B. Incumbent Emulation and Sensing
D. Use Case: Detection-to-Mitigation
The managed incumbent is an emulated FSS earth-station receiver. Because the USRP B210 cannot operate at FR3, a FieldFox N9953B serves as the physical sensing device, while the victim-receiver characteristics are applied analytically to the calibrated captures. The measured victim-band power is mapped from the IF through the verified frequency plan, referred to the victim plane using the calibrated Vivaldi antenna and RF chain, and weighted according to the International Telecommunication Union Radiocommunication Sector (ITUR) S.1428 off-axis gain [16]. With N = kTsys B, where k is the Boltzmann constant, Tsys is the system noise temperature, and B is the reference bandwidth, the ITU-R S.1432 criterion max I/N ≤ −10 dB [17] defines the interference budget IFSS enforced by the policy gate. FieldFox captures are timestamped
The enforcement workflow consists of five stages. First, during admission, the Policy Interpreter converts the experiment request into a spectrum grant specifying its frequency, bandwidth, and power constraints. The Experiment Steward then deploys the experiment and verifies its configuration using calibrated spectrum measurements. Second, a controlled drift is introduced by switching the experiment to full-power, full-bandwidth operation. Due to the spectrum inversion, the resulting emission skirt increases interference in the lower-adjacent FSS sub-band and eventually exceeds the protection budget. Third, the Spectrum Sentinel performs periodic FieldFox captures and maps the measured IF power to the virtual FSS victim reference plane. It detects the violation and correlates
102
Value
NR numerology / PRBs DL/UL pattern gNB TX-power cap Pmax B210 TX gain range B210 max RX input IF carrier (FR1) FR3 LO / injection FR3 carrier RF interconnect
µ = 1 (30 kHz SCS) / 106 DDDDDDDSUU 20 dBm 0–89.75 dB −15 dBm 3.58 GHz 13.58 GHz / high-side 10 GHz conducted IF chains (attenuators, splitters, BPF/LPF); OTA (2 m Tx/Rx) 100 ms ITU-R S.1432 I/N −10 dB virtual (FieldFox + S.1428/S.1432) Tsys = 150 K / B = 36 MHz UDP DL iperf3 Claude Sonnet 5 10
Control interval ∆t Protection criterion Victim realization Victim noise / ref. bandwidth Traffic Agent runtime Trials per condition Ntrial
the observed spectrum with the experiment configuration. The lower-adjacent emission provides evidence of the inverted-skirt mechanism. Fourth, the Zone Orchestrator receives the attribution, candidate mitigations, and their projected post-action interference and disruption cost, then selects the lowest-cost candidate that meets the protection target. The selected action is a PRB-region mask that removes the PRBs mapped to the protected FSS subband while retaining the remaining resources. The proposed action is then submitted to the policy gate, which independently computes its expected outcome before admitting it. Finally, the Experiment Steward applies the mask through the O-RAN control plane, and the Spectrum Sentinel performs an independent post-action measurement to verify that the interference budget is restored. If successful, the experiment continues with the reduced allocation rather than being terminated. In contrast, the procedural baseline shuts down the experiment upon the same violation. Throughout the experiment, the reflex independently monitors the budget and can immediately apply the pre-computed PRB mask, providing a deterministic protection path independent of agent execution. V. N UMERICAL R ESULTS We evaluate A-RDZ over Ntrial = 10 drift trials and compare it with a procedural shutdown baseline. All enforcement stages operate on the hardware-in-the-loop path: FieldFox measurements provide the interference estimate, which is referred to the virtual victim plane using the calibration of Sec. IV. Postactuation verification uses an independent capture. All timestamps are wall-clock measurements from a single host. The key experimental parameters are summarized in Table II. A. Detection-to-Mitigation Timeline Figure 3 compares the latency of the agentic enforcement path with the agent-independent reflex. The reflex deadline of 200 ms represents the design sense-to-actuate budget of two 100 ms control intervals. With the agent, the enforcement loop includes detection, decision, actuation, and verification. The
reflex deadline (200 ms) 10−1
control interval (∆t=100 ms)
10−4 Detect
Decide
Actuate
Verify
Reflex compute (no agent)
Fig. 3. Latency decomposition of the agentic enforcement loop over Ntrial = 10 trials. drift (full buffer) 10
Margin to budget [dB]
Parameter
Latency [s]
TABLE II E XPERIMENT PARAMETERS .
mask applied
detect
verified
5 0
max IFSS (I/N ≤ -10 dB, ITU-R S.1432) A-RDZ: masked, exp. continues (cost 0.24) baseline: shutdown (cost 1.00)
−5 −10 0
20
40
60
Time [s] Fig. 4. Interference margin, estimated at the virtual FSS victim reference plane, during one representative trial.
median end-to-end latency is 73.74 s, obtained as the sum of the four per-stage medians plotted in Fig. 3. In contrast, the reflex decision itself requires only 0.14 ms of computation on average because it applies a pre-computed protective action without LLM-based reasoning or agent-mediated tool execution. The end-to-end protection response remains bounded by the 100 ms sensing/control interval and the latency of the sensing path. This difference highlights the complementary roles of the two paths: the reflex provides fast, deterministic protection, whereas the agentic path performs higher-level diagnosis, mitigation selection, and post-action verification. B. Protection Compliance and Experiment Utility We define the instantaneous margin relative to the protection max budget as Margin(t) = 10 log10 (IFSS /I(t)) [dB], so that a positive margin denotes compliance and a negative margin denotes a max violation of IFSS . Figure 4 shows a representative drift episode. max At the detected violation, I/IFSS = 6.75, corresponding to a margin of −8.29 dB. The agent selects a PRB-region mask, max reducing the interference to I/IFSS = 0.496 and restoring a positive margin of 3.05 dB. We also define the utility by a mitigated retained allocation P P grant p as η = ) , the ratio of n e mn (pn ) / n e mn (pn achievable downlink rate to that of the granted operating point, where e(·) is the TS 38.214 spectral-efficiency mapping [19] and mn denotes the MCS index selected for PRB n under allocation pn . The disruption cost is C = 1−η, so a full shutdown gives C = 1. Across all 10 trials, the violation is successfully detected and mitigated, and the experiment retains 76.04% of its utility (η = 0.76). In comparison, the shutdown baseline terminates the experiment, corresponding to a normalized cost of C = 1.00 versus C = 0.24 for A-RDZ.
VI. R ELATED W ORK RDZ concept and prototypes. The RDZ concept, its stakeholder analysis, and its policy/experiment/spectrum decomposition were articulated in [3]. Prototype zone management systems include Powder-RDZ and its OpenZMS zone management system [4], [5], the COSMOS ZMS [6], and reactive spectrum-sharing extensions on POWDER [20]. FlexRDZ [7] is closest in spirit to the autonomy, managing mobile transmitters with hierarchical task networks and digital-twin prediction; its planning, however, remains procedural over a fixed task vocabulary, whereas the A-RDZ reasons from heterogeneous evidence while enforcing policy constraints. Agentic AI for wireless. G ENESIS [11] introduced multiagent autonomous R&D on production O-RAN testbeds and is the substrate we extend; AgentRAN [21] proposes an agentic architecture for autonomous control of open 6G networks, and ALLSTaR [22] uses LLMs to synthesize and test Medium Access Control (MAC) schedulers from intent. To our knowledge, no prior work has applied agentic systems to the operation of a spectrum-sharing zone, where the object of control is not a network’s own performance but the protection relationship between an experiment and an incumbent. FR3 coexistence. Upper mid-band NR–FSS coexistence has been studied predominantly through normative emission masks and statistical propagation models [23], [24], with FR3 channel measurements now emerging from real hardware [25], [26]. Our work complements these efforts by demonstrating agentic zone management over a hardware-in-the-loop FR3 coexistence experiment. VII. C ONCLUSIONS We introduced A-RDZ, an agentic realization of the RDZ architecture in which LLM-based agents perform spectrum, experiment, and policy management while deterministic gates and near-RT reflexes preserve incumbent protection. We demonstrated the architecture end-to-end on a hardware-in-the-loop FR3 O-RAN testbed with a 5G NR experiment and an emulated FSS incumbent. The results show that agentic operation can detect and mitigate unexpected interference while retaining experiment utility, but its non-deterministic latency motivates separating adaptive reasoning from safety-critical protection. Future work will address multi-experiment coordination, boundary sensing, formal verification of the agent–gate interface, and larger-scale RDZ deployment. R EFERENCES [1] A. Bazzi, R. Bomfin et al., “Upper mid-band spectrum for 6G: Vision, opportunity and challenges,” IEEE Commun. Mag., vol. 64, no. 1, pp. 206–212, Jan. 2026. [2] Z. Cui, P. Zhang, and S. Pollin, “6G wireless communications in 7–24 GHz band: Opportunities, techniques, and challenges,” in Proc. IEEE Int. Symp. Dyn. Spectr. Access Netw. (DySPAN), May 2025, pp. 1–8. [3] M. Zheleva, C. R. Anderson et al., “Radio dynamic zones: Motivations, challenges, and opportunities to catalyze spectrum coexistence,” IEEE Commun. Mag., vol. 61, no. 6, pp. 156–162, Jun. 2023. [4] D. Johnson, D. Maas et al., “Building radio dynamic zones with OpenZMS,” IEEE Trans. Cogn. Commun. Netw., vol. 11, no. 3, pp. 1410– 1426, Jun. 2025.
[5] ——, “PowDER-RDZ: Prototyping a radio dynamic zone using the POWDER platform,” in Proc. IEEE Int. Symp. Dyn. Spectr. Access Netw. (DySPAN), Washington, DC, USA, May 2024. [6] C. E. C. Bastidas, J. N. Doshi et al., “COSMOS ZMS: A radio dynamic zone management system using the COSMOS platform,” in Proc. IEEE Int. Symp. Dyn. Spectr. Access Netw. (DySPAN), May 2026, pp. 184–190. [7] A. Gottipati and J. Van der Merwe, “FlexRDZ: Autonomous mobility management for radio dynamic zones,” in Proc. IEEE Future Netw. World Forum (FNWF), Baltimore, MD, USA, Nov. 2023. [8] S. Yao, J. Zhao et al., “ReAct: Synergizing reasoning and acting in language models,” in Proc. 11th Int. Conf. Learn. Represent. (ICLR), Mar. 2023, pp. 1–33. [9] N. Shinn, F. Cassano et al., “Reflexion: Language agents with verbal reinforcement learning,” in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), Dec. 2023, pp. 8634–8652. [10] Q. Wu, G. Bansal et al., “AutoGen: Enabling next-gen LLM applications via multi-agent conversations,” in Proc. First Conf. Lang. Model. (COLM), Aug. 2024, pp. 1–43. [11] T. Aghayev, M. Elkael et al., “GENESIS: Harnessing AI agents for autonomous 6G RAN synthesis, research, and testing,” 2026. [Online]. Available: https://arxiv.org/abs/2605.27360 [12] J. P. Anderson, “Computer security technology planning study,” Air Force Electronic Systems Division, Tech. Rep. ESD-TR-73-51, Oct. 1972. [13] M. Alshiekh, R. Bloem et al., “Safe reinforcement learning via shielding,” in Proc. AAAI Conf. Artif. Intell. (AAAI), Feb. 2018, pp. 2669–2678. [14] N. Nikaein, M. K. Marina et al., “OpenAirInterface: A flexible platform for 5G research,” ACM SIGCOMM Comput. Commun. Rev., vol. 44, no. 5, pp. 33–38, Oct. 2014. [15] M. Mezzavilla, A. Rasteh et al., “6G prototyping in the upper mid-band (7–24 GHz),” in Proc. IEEE Wireless Commun. Netw. Conf. (WCNC), Mar. 2025, pp. 1–3. [16] ITU-R, “Reference FSS earth-station radiation patterns for use in interference assessment involving non-GSO satellites in frequency bands between 10.7 GHz and 30 GHz,” International Telecommunication Union, Radiocommunication Sector, Tech. Rep. Recommendation ITUR S.1428-1, Aug. 2001. [Online]. Available: https://www.itu.int/dms pubrec/itu-r/rec/s/R-REC-S.1428-1-200102-I!!PDF-E.pdf [17] ——, “Apportionment of the allowable error performance degradations to FSS hypothetical reference digital paths arising from timeinvariant interference,” International Telecommunication Union, Radiocommunication Sector, Tech. Rep. Recommendation ITU-R S.1432-1, Jun. 2006. [Online]. Available: https://www.itu.int/dms pubrec/itu-r/rec/s/R-REC-S.1432-1-200604-I!!PDF-E.pdf [18] R. Schmidt, M. Irazabal, and N. Nikaein, “FlexRIC: An SDK for nextgeneration SD-RANs,” in Proc. ACM CoNEXT, Virtual Event / Munich, Germany, Dec. 2021, pp. 411–425. [19] 3GPP, “5G; NR; Physical Layer Procedures for Data, version 16.2.0,” 3rd Gener. Partnership Project (3GPP), Tech. Rep. TS. 38.214, Jul. 2020. [20] A. Sarbhai, D. Johnson et al., “Reactive interference management for radio astronomy in radio dynamic zones using ASTRA,” in Proc. IEEE Int. Symp. Dyn. Spectr. Access Netw. (DySPAN), May 2025, pp. 1–10. [21] M. Elkael, S. D’Oro et al., “AgentRAN: An agentic AI architecture for autonomous control of open 6G networks,” IEEE Commun. Mag., 2026, early access. [22] M. Elkael, M. Polese et al., “ALLSTaR: Automated LLM-driven scheduler generation and testing for intent-based RAN,” IEEE Trans. Mobile Comput., 2026, early access. [23] ITU-R, “Modelling and simulation of IMT networks and systems for use in sharing and compatibility studies,” International Telecommunication Union, Radiocommunication Sector, Tech. Rep. Recommendation ITU-R M.2101-0, Feb. 2017. [Online]. Available: https://www.itu.int/dms pubrec/itu-r/rec/m/R-REC-M.2101-0-201702-I!!PDF-E.pdf [24] B. Lim and M. Vu, “Interference analysis for coexistence of terrestrial networks with satellite services,” IEEE Trans. Wireless Commun., vol. 23, no. 4, pp. 3146–3161, Apr. 2024. [25] D. Shakya, M. Ying et al., “Urban outdoor propagation measurements and channel models at 6.75 GHz FR1(C) and 16.95 GHz FR3 upper midband spectrum for 5G and 6G,” in Proc. IEEE Int. Conf. Commun. (ICC), Montreal, Canada, Jun. 2025, pp. 3291–3296. [26] ——, “Comprehensive FR1 (C) and FR3 lower and upper mid-band propagation and material penetration loss measurements and channel models in indoor environment for 5G and 6G,” IEEE Open J. Commun. Soc., vol. 5, pp. 5192–5210, Jul. 2024.