Demonstrates on live O-RAN hardware that two independently correct rApps jointly produce recurring unsafe oscillations in shared radio-resource partitions.
Taming the Agentic RAN (arXiv:2609.18857) demonstrates that two independently correct rApps—one protecting latency SLAs and another maximizing energy efficiency—jointly drive recurring opposing excursions of shared radio resource partitions (PRB quotas) on a live OpenAirInterface (OAI) testbed. This instability arises from delayed feedback between opposing best-response loops, causing limit cycles that neither agent produces alone. To resolve this, the authors propose AURA, a lightweight arbitration layer that enforces three conditions on agent proposals: feasibility invariants (quota caps), dwell times exceeding control loop delays, and a deadband for negligible changes. AURA reduces shared-state excursions by an order of magnitude (from 8.4 to 0.4 PRB amplitude) and virtually eliminates cross-slice throughput starvation, though it does not improve the protected slice’s own latency compliance.
The material addresses a core reliability challenge in agentic O-RAN, where autonomous AI agents—implemented as rApps in the near-real-time RIC—make distributed decisions over shared radio resources. Its central empirical finding is that two rApps can each be locally correct by their own control objectives, yet their interaction over shared radio-resource partitions can produce recurring unsafe oscillations on live O-RAN hardware. In other words, the failure is not caused by a defective individual agent, but by emergent instability arising from concurrent, independently optimized control loops acting on the same RAN state.
A key contribution is the identification and demonstration of this interaction-level failure mode in a realistic O-RAN setting, rather than only in simulation or abstraction. The work also introduces a stability-guaranteed arbitration mechanism that mediates rApp decisions before they are applied to the RAN. Instead of allowing each agent to act directly on shared partitions, the arbitration layer reconciles competing requests and constrains admissible actions so that the resulting closed-loop behavior avoids unsafe oscillatory dynamics. The insight is that agentic RAN systems need an interaction-control layer: per-rApp correctness is necessary but not sufficient for system-level safety.
This matters because O-RAN’s open, programmable architecture is intended to enable multiple vendors and AI-driven rApps to coexist and improve network operation. But that flexibility creates an assurance gap: independently developed agents can interfere in ways that are hard to predict from their individual specifications. By showing that safe behavior must be enforced at the system level, and by providing an arbitration approach with stability guarantees, the material offers a practical path toward deploying autonomous RAN agents with stronger reliability, safety, and operational predictability.