PaperlayerView exact revision on arXiv

Source-linked AI summary

Taming the Agentic RAN: Stability-Guaranteed Arbitration of Autonomous AI Agents in O-RAN

Seyed Bagher Hashemi Natanzi, Bo Tang

arXiv:2609.18857v1cs.NIcs.AIcs.SYeess.SY

TL;DR

Independently acting O-RAN agents can drive recurring opposing excursions in shared radio resources, while existing conflict mitigation assumes statically known applications. AURA arbitrates declared actions under feasibility, dwell-time, and deadband checks; on a live O-RAN testbed it sharply reduces cross-slice starvation, but does not improve the protected slice’s latency compliance.

  • Problem

    Existing conflict-mitigation methods assume statically known applications, leaving runtime conflicts from autonomous agents with opaque or evolving behavior unaddressed.

    Source: §I · p. 1

  • Method

    AURA admits agent proposals only when feasibility invariants, per-variable dwell times, and a deadband are satisfied, with priorities resolving conflicts.

    Source: §I · p. 1, §III · p. 3

  • Results

    40–55% slice-2 throughput violations under Direct and Single fell to 0.3% under AURA, while slice-1’s latency-violation rate did not improve.

    Source: §VII · p. 8

  • Takeaways & Limitations

    AURA enforces a feasible operating point and makes explicit the system-wide trade-off between cross-slice starvation and the protected slice’s latency compliance.

    Source: §VII · p. 8, §VIII · p. 8

    The empirical demonstration is scoped to one cell, two structurally opposed agents, and RF-simulated radio.

    Source: §VII · p. 8

Abstract

from arXiv · show

The O-RAN control plane is becoming agentic: autonomous AI agents, deployed as rApps by different vendors, independently close control loops over shared radio resources. We demonstrate on a live O-RAN system that this independence is unsafe. Two agents with individually correct objectives, one protecting a latency SLA and one maximizing utilization for energy efficiency, jointly drive recurring opposing excursions of the shared resource partition that neither produces alone. Existing conflict-mitigation mechanisms presume a statically known application population and cannot govern agents whose behavior emerges at run time. We present AURA, a lightweight arbitration layer that admits agent actions only when they satisfy feasibility invariants, per-variable dwell times, and a deadband, and we prove the arbitrated system converges to a feasible operating point. Implemented on an OpenAirInterface (OAI) testbed with measured one-way latency and throughput, AURA reduces recurring shared-state excursions by more than an order of magnitude (from 8.4 to 0.4 PRB amplitude) and virtually eliminates cross-slice throughput starvation (from 40-55% to 0.3%), while leaving the protected slice's own latency compliance unchanged, a trade-off the convergence guarantee makes explicit.

I. Introduction

Autonomous agents with individually reasonable objectives can destabilize shared O-RAN resources through delayed, opposing control loops. AURA addresses this runtime conflict with proposal-based arbitration, while existing mitigation assumes known application populations.

I. Introduction

  • Individually reasonable agents can repeatedly drive opposing excursions of a shared resource partition, a pathology absent when either agent runs alone.The SLA agent expands its quota while the energy agent interprets spare capacity as waste and reduces the total budget, repeating the cycle after disturbances.Source: §I · p. 1

II. Related Work

  • Existing conflict-mitigation methods address xApps with statically known parameter footprints, whereas agentic conflicts emerge at run time from opaque or evolving behavior.AURA inspects declared effects on shared state rather than application internals.Source: §I · p. 1, §II · p. 2

I. Introduction

  • AURA admits declared agent actions only when feasibility, dwell-time, and deadband conditions hold, resolving conflicts by priority and proving convergence to a feasible operating point.It does not require knowledge of agents’ decision logic and is validated on a live O-RAN testbed.Source: §I · pp. 1–2
  • The study’s contributions include a formal delayed opposing best-response model, a limit-cycle proof, and an implementation of per-slice PRB quota enforcement in the OAI NR MAC downlink pre-processor.It also measures the full control-loop latency budget from agent decision to MAC enforcement.Source: §I · p. 2
  • AURA reduces shared-state excursions by an order of magnitude and virtually eliminates coexisting-slice throughput starvation, without improving the protected slice’s latency compliance.The paper presents this as a system-wide trade-off rather than per-agent optimality.Source: §I · p. 2

II. Related Work

  • The authors position AURA’s contribution as applying established stability tools to O-RAN and validating both the instability and its remedy on a real 5G system.Their agents are rule-based in this study so the observed behavior can be attributed to interaction structure rather than policy sophistication.Source: §II · p. 2

III. System Model

The system model represents a single-cell RAN through shared slice quotas and a total resource cap, with latency and throughput SLAs observed by two independent agents. AURA arbitrates their state-changing proposals using feasibility, dwell-time, and deadband checks.

III. System Model

  • The shared state is the slice-quota vector and total cap, with the sum of quotas constrained by the cap and quotas enforced in the MAC downlink pre-processor.When their sum exceeds the cap, the enforcement mechanism scales quotas down.Source: §III · p. 2
  • Slice 1 has a one-way-delay target and slice 2 a receiver-throughput target; for fixed offered load, their metrics improve monotonically with their respective effective service rates.Utilization is measured relative to the cap, so tightening the cap can raise utilization without serving additional traffic.Source: §III · p. 2
  • The SLA agent adjusts slice quotas from observed delay and throughput, while the efficiency agent adjusts the cap from observed utilization; neither observes the other’s objectives or proposals.The agents are rApps acting on multi-second horizons, and their rule-based implementation is not required by the arbitration design.Source: §III · p. 3
  • An observation at time t reflects telemetry from an earlier interval, with loop delay defined as τloop = To + d.Here d aggregates measurement, transport, and enforcement latency.Source: §III · p. 3
  • In the arbitrated regime, each proposal declares its agent, shared-state effect, priority, and TTL, and is admitted only if the resulting state satisfies feasibility invariants.A variable must also clear its dwell time and the proposed change must exceed the deadband; conflicts are resolved by priority.Source: §III · p. 3

IV. Instability and Its Removal

The model links opposing best responses to a shared fixed point where both constraints bind. Their composition is increasing, providing the structural basis for analyzing the agents’ interaction.

A. Why the agents fight

  • Under saturating background load, a tighter cap requires a larger protected-slice quota to maintain latency, while a larger quota lowers utilization and drives the energy agent to reduce the cap.The model defines best-response maps for these coupled actions.Source: A. Why the agents fight · p. 3
  • Because both best-response maps are decreasing, their composition is increasing.Source: A. Why the agents fight · p. 3
  • The fixed point x* is the operating point at which both constraints bind.Source: A. Why the agents fight · p. 3

B. Uncoordinated dynamics oscillate

With delayed observations and gain above one, direct opposing updates make the fixed point unstable and produce a limit cycle. The experiment tests whether excursions recur across independent load cycles and disappear when either agent is disabled.

B. Uncoordinated dynamics oscillate

  • When local gain exceeds one and delayed agents act every Ta with τloop ≥ Ta, the direct-regime fixed point is unstable and trajectories enter a limit cycle.The cycle amplitude is bounded below by δ · ⌈τloop/Ta⌉.Source: B. Uncoordinated dynamics oscillate · p. 3
  • Stale observations let an agent continue stepping before its prior actions appear in telemetry; gain above one amplifies the resulting overshoot, while bounded state confines trajectories to a finite lattice.The resulting deterministic dynamics eventually become periodic, with period greater than one because the fixed point is unstable.Source: B. Uncoordinated dynamics oscillate · p. 3
  • 3 of 5 Direct repetitions showed cap movement during the second independent load cycle, with amplitudes 0, 2, 0, 2, and 6 PRBs; amplitude was zero when either agent ran alone.This tests the predicted recurrence under disturbance and disappearance with either agent disabled.Source: B. Uncoordinated dynamics oscillate · p. 3

C. Arbitration restores convergence

AURA’s dwell time, deadband, and priority-ordered admission make arbitrated actions converge to a feasible state while preserving invariants at every step. The architecture places a central arbiter between agents and the network.

C. Arbitration restores convergence

  • With dwell time τd > τloop and deadband 0 < ϵ ≤ min(δ, δc), arbitration terminates within finitely many admitted actions at a feasible point within ϵ, with q1 ≥ qfloor maintained.The guarantee does not claim the terminal point is optimal.Source: C. Arbitration restores convergence · p. 3
  • Dwell time serializes modifications against fresh telemetry, while priority-ordered nonredundant actions improve a finite state lattice and admission checks preserve invariants.This prevents stale-observation overshoot and yields finite termination.Source: C. Arbitration restores convergence · pp. 3–4
  • The guarantee ensures feasibility and SLA-floor safety, not optimality or improved compliance for the protected slice beyond its pre-existing allocation.AURA’s measured trade-off is damping shared-state excursions and protecting the coexisting slice without improving the protected slice’s own SLA compliance.Source: C. Arbitration restores convergence · p. 4

A. Architecture

  • The arbiter interposes on the agents’ action path, updates the quota table, and receives telemetry through the E2 MAC service model into a store shared with agents and measurement.The measurement harness also reads the shared telemetry store; events are logged on a single host-clock timeline.Source: A. Architecture · p. 4

B. PRB quota enforcement in the OAI NR MAC

The implementation enforces per-slice PRB quotas in the OAI MAC and routes agent proposals through an arbiter. Measurements use a dedicated one-way-delay probe and receiver-side throughput counters in a controlled OAI/FlexRIC setup.

B. PRB quota enforcement in the OAI NR MAC

  • The OAI downlink pre-processor enforces per-slice PRB quotas using a table mapping slice identifiers to PRB limits and UEs to slices; the cap scales quotas when their sum exceeds C.The implementation replaces the upstream E2 slice service model emulator with callbacks bound to the same quota table.Source: B. PRB quota enforcement in the OAI NR… · p. 4

C. The arbiter

  • The arbiter tracks shared state, per-variable modification times, and invariants, then admits or rejects proposals and returns the violated condition to rejected agents.It requires no model of agent internals, so rule-based and learning-based agents use the same proposal schema.Source: C. The arbiter · p. 4

D. Measuring what the agents observe

  • One-way latency is measured with a timestamped UDP probe at the SLA-protected UE, while throughput comes from receiver-side UE counters.The probe uses the shared host clock in the RF-simulated deployment; an echo test verifies bidirectional reachability before each run.Source: D. Measuring what the agents observe · p. 4

A. Setup

  • Experiments use an OAI core and RF-simulated gNB with two UEs and FlexRIC; four regimes compare fixed quotas, one agent, direct two-agent control, and AURA arbitration.Runs use two load cycles, with five repetitions per regime; UDP avoids congestion-control confounds.Source: A. Setup · pp. 4–5

B. Actuator validation

Quota validation shows that the MAC actuator controls delivered throughput and shifts throughput directionally under two-UE quota changes in the validated load region. The tests also identify a dual-softmodem compute limitation at higher loads.

B. Actuator validation

  • Quota tests drive the pre-processor directly, isolating MAC enforcement from the E2 control path, and verify bidirectional UE-path reachability before experiments.The independent echo check caught an uplink routing artifact while downlink remained unaffected.Source: B. Actuator validation · pp. 5–6
  • Raising one UE’s quota from 13 to 38 PRBs increased receiver goodput from 17.0 to 43.1 Mbps, a 2.54× increase; MAC allocation tracked the configured quota exactly.This single-UE saturating-UDP sweep isolates the quota actuator from dual-UE scheduling interactions.Source: B. Actuator validation · p. 5
  • At 22 Mbps offered per UE, equal quotas yielded 23.1/23.1 Mbps; at 28 Mbps offered, a 13:38 quota split yielded 19.1/29.4 Mbps.These results show fairness under equal quotas and directional throughput shifts under asymmetric quotas at moderate offered load.Source: B. Actuator validation · p. 5
  • In the representative two-cycle run, Direct continues moving C in the second cycle, while AURA keeps q1 essentially frozen and damps C; measured ℓ1 remains noisy under both.The figure’s regime distinction is in the shared-state trajectories, not a clean separation in measured latency.Source: Figure 2 · p. 6
  • Near the dual-UE ceiling, one UE can starve from RFsim shared software-PHY compute load even while MAC-level quota assignment tracks configured values exactly.Agent experiments therefore use offered loads within the validated fair/directional region.Source: B. Actuator validation · p. 6

C. Uncoordinated agents destabilize what each protects alone

Individually reasonable agents destabilize the shared resource cap when they act together, while AURA sharply reduces that movement and cross-slice starvation. Its trade-off is that slice-1 latency compliance remains unimproved.

C. Uncoordinated agents destabilize what each protects alone

  • With Asl alone, q1 moved 19.0 PRBs while C was untouched; with both agents, C moved 8.4 settled PRBs and 2.0 PRBs in the second cycle.In 3 of 5 repetitions, C moved at least 2 PRBs during the second cycle, supporting recurring interaction rather than a one-time transient.Source: C. Uncoordinated agents destabilize… · p. 6, Table I · p. 6
  • Slice-2 violations fell from 40–55% in Single and Direct to 0.3% under AURA (Static: 0.6%), while slice-1 violations remained 80–93% across regimes.Baseline one-way delay frequently exceeds 18 ms; a static control at AURA’s operating point had 87.0% slice-1 violations (±2.9), with precise attribution left for future work.Source: C. Uncoordinated agents destabilize… · pp. 6–7, Table II · p. 6

D. Arbitration removes the cross-slice pathology, at an explicit cost

  • Under AURA, settled C amplitude falls from 8.4 to 0.4 PRBs, and second-cycle movement falls from 2.0 to 0.4 PRBs.AURA also reduces q1’s settled amplitude from 10.4 to 1.4 PRBs.Source: Table I · p. 6, D. Arbitration removes the cross-slice… · p. 7
  • AURA rejected 548 proposals: 537 (98%) for violating the feasibility invariant, 11 (2%) for dwell, and none for deadband.The deadband check did not fire because the fixed step size δ = 2 never fell below the configured ϵ = 1.Source: D. Arbitration removes the cross-slice… · p. 7
  • AURA leaves slice-1 latency violations unimproved, trading that metric for near-complete coexisting-slice protection and reduced shared-state movement.The single-variable proposal schema cannot express an atomic action that grows both q1 and C.Source: D. Arbitration removes the cross-slice… · p. 7

E. Control-loop latency budget

AURA’s dwell time is set above the measured end-to-end control-loop delay, which is dominated by enforcement delivery in this deployment. The reported evaluation fixes the agent period and leaves delay, step-size, and population scaling for future work.

E. Control-loop latency budget

  • A 264.7 ms median enforcement-delivery latency, combined with a 5 s agent period and ~1 s observation overhead, yields an estimated 7 s loop delay and an 8 s dwell time.Arbitrated runs are refused unless the configured dwell time exceeds the estimated loop delay.Source: E. Control-loop latency budget · p. 7
  • Arbiter computation is negligible; file-driven enforcement, not agent reasoning or arbitration logic, determines the safe dwell-time floor in this deployment.A production E2-based path may differ substantially and would require remeasurement.Source: E. Control-loop latency budget · p. 8

F. Sensitivity

  • The evaluation fixes the agent period at 5 s and leaves agent-period, step-size, and agent-population scaling untested.The authors do not claim platform-scale concurrent-agent counts.Source: F. Sensitivity · p. 8

VII. Discussion and Future Work

AURA makes a safety–performance trade-off explicit: it bounds shared state and reduces cross-slice starvation, without improving the protected slice’s latency compliance. The discussion identifies limits of the current atomic proposals and evaluation scope, while outlining broader applicability and future extensions.

VII. Discussion and Future Work

  • Single-variable proposals cannot jointly increase slice-1 quota and the energy cap; composite proposals could request both atomically while preserving post-state feasibility.This extension requires broadening the agents’ action space, not changing the arbitration architecture.Source: §VII · p. 8
  • The evaluation uses one cell, two opposed agents, and RF simulation; multi-cell, many-agent OTA tests and learned-policy validation remain future work.Conflict-rate scaling with agent population remains open.Source: §VII · p. 8
  • The formal instability template applies beyond slicing when controllers have opposing monotone responses over shared variables and delayed feedback, including power-control/interference and traffic-steering/energy settings.The paper also proposes verifying declared effects against observed outcomes to quarantine diverging agents.Source: §VII · p. 8

VIII. Conclusion

  • AURA reduces shared-state movement by an order of magnitude and cuts slice-2 throughput violations from 40–55% under Direct and Single to 0.3%, without improving slice-1 latency compliance.The reported slice-1 violation rate is 92.9% under AURA; Direct’s 84.5% rate comes with slice-2 throughput starvation.Source: §VIII · p. 8, §VII · p. 8
  • AURA evaluates proposed state modifications rather than internal agent logic, and the authors plan to release the enforcement patch, arbiter, and measurement harness.The conclusion frames the arbiter as an enforcement boundary between agent actions and the network.Source: §VIII · p. 8

Appendix

The appendix reports CPU utilization across regimes and pinned core sets, with values remaining near 10%.

Appendix

  • CPU utilization was sampled from host /proc/stat every 5 seconds during the agent matrix.The table pools five repetitions per regime.Source: Appendix · p. 8, Table III · p. 9
  • CPU means stay near 10%, with gNB 95th percentiles under 13.4%, well below the 64–71% dual-UE starvation band.The largest regime difference is roughly 2 percentage points, between Static and AURA on the gNB set.Source: Appendix · p. 9, Table III · p. 9
  • Table III reports mean CPU utilization by regime for pinned gNB, UE1, and UE2 core sets.The caption identifies gNB cores 0–7, UE1 cores 8–11, and UE2 cores 12–15.Source: Table III · p. 9
Loading 2609.18857v1…