Conflict-Aware Autonomous Resource Orchestration for 5G Edge Slicing Using Multi-Agent Systems and LLMs

Conflict-Aware Autonomous Resource Orchestration for 5G Edge Slicing Using Multi-Agent Systems and LLMs Nathan R. Prediger1 , Rodrigo Moreira2 , Juliano Araujo Wickboldt1 , Tereza C. Carvalho3 , Pedro Oliveira4 , Paulo Novais4 , Flávio de Oliveira Silva4

arXiv:2610.14903v1 [cs.NI] 10 Oct 2026

1

Federal University of Rio Grande do Sul (UFRGS), Porto Alegre, Brazil 2 Federal University of Viçosa (UFV), Minas Gerais, Brazil 3 University of São Paulo (USP), São Paulo, Brazil 4 University of Minho (UMinho), Braga, Portugal Emails: nrprediger@inf.ufrgs.br, rodrigo@ufv.br, jwickboldt@inf.ufrgs.br, terezacarvalho@usp.br, pedro.jose.oliveira@algoritmi.uminho.pt,{pjon, flavio}@di.uminho.pt

operating at the network edge (Edge Computing), where strict latency bounds and reliable access to bandwidth resources are essential to preserve Quality of Service (QoS). In scenarios such as remote medical procedures or industrial control systems, Network Slicing creates isolated logical networks over shared infrastructure and enables differentiated resource allocation, ensuring that sensitive applications retain reserved or prioritized access to resources even under high demand or noisy neighbors [1], [2]. Orchestrating these slices at the edge in a distributed scenario poses more challenges than in the network core [3]. Unlike the core, the edge has scarce and highly contended computational resources (i.e., CPU, memory, and bandwidth). Centralized management approaches can face computational and responsiveness challenges when multiple resource-contention events occur simultaneously [4]. A second, more subtle problem compounds this: even a system that reacts quickly to contention still needs to know what to prioritize. Today, the mapping from a high-level operator intent – a live event that demands more bandwidth, a maintenance window, or an emergency response – to the concrete resource weight that a slice should receive is largely manual, performed by operators, or hard-coded into static rules that cannot adapt to intents unforeseen at design time. In this context, Multi-Agent Systems (MAS) emerge as a promising approach to the first problem. The MAS paradigm distributes the network’s decision-making: each slice is represented by an autonomous agent capable of participating I. I NTRODUCTION in resource negotiation and contention resolution through The evolution of 5th Generation Mobile Network (5G) net- coordinated interactions with other agents. This approach works has intensified the need for mission-critical applications aligns with the Self-Organizing Networks concept, enabling distributed and adaptive responses to abrupt load variations. © 2026 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing Yet auction-based MAS coordination still depends on operatorthis material for advertising or promotional purposes, creating new collective works, for defined economic parameters and priorities. Without an adaptive resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. mechanism for updating these parameters, they can remain as Accepted author manuscript for publication in the Proceedings of the 2027 IEEE static as the heuristics they replace, effectively shifting the Consumer Communications & Networking Conference (CCNC 2027), Las Vegas, Nevada, USA, January 9–12, 2027. intent-to-infrastructure gap one layer up the control stack. Abstract—5G edge network slicing must reconcile two coupled requirements: understanding high-level operator intent and reallocating scarce computing and bandwidth resources under dynamic contention. Centralized optimization can face computational and responsiveness challenges under rapidly changing edge contention, and Multi-Agent System (MAS)-based auctions can support resource arbitration but still depend on operator-defined economic parameters and priorities. Embedding a Large Language Model (LLM) directly in a recurring control loop, by contrast, may introduce non-negligible inference latency and non-deterministic behavior. This paper proposes a two-timescale architecture that addresses this gap: an LLM-based Strategist translates naturallanguage network intents into economic policy parameters (budget, income, and bidding-aggressiveness) at the macro scale, while a deterministic, Vickrey-inspired second-price mechanism run by SPADE-based slice agents resolves recurring resource contention at the operational scale, with the resulting allocations enforced by patching Kubernetes CPU limits and Calico CNI bandwidth annotations directly on User Plane Function pods. We evaluate the architecture end-to-end on a real Kubernetes/Open5GS edge cluster across two complementary scenarios, each repeated 30 times: a Smart-City scenario, in which natural-language publicsafety intents propagate into measurable changes in slice bids and bandwidth allocations within a few auction rounds, and a Hierarchical-slicing scenario, in which the mechanism provides priority-dependent resource and QoE differentiation, adapts to abrupt demand changes, and preserves configured minimum guarantees while maintaining near-complete allocation of the shared bandwidth pool. Index Terms—5G, Network Slicing, Edge Computing, MultiAgent Systems, Large Language Models, Intent-Based Networking, Resource Orchestration

This paper addresses this gap through a two-timescale architecture that separates semantic intent interpretation from deterministic resource arbitration. At the macro scale, a Large Language Model (LLM) acts as a Strategist, interpreting naturallanguage network intents and translating them into economic policy parameters, including slice budget, income rate, and bidding-aggressiveness weight. At the operational scale, these policies parameterize a deterministic Vickrey-inspired secondprice mechanism executed by the MAS, in which slice agents compete for available resources according to their economic parameters and observed resource conditions. The resulting allocations are enforced directly on the infrastructure by updating CPU limits and Calico CNI bandwidth annotations on the User Plane Function (UPF) pods of a real Kubernetes/Open5GS edge cluster, without requiring manual intervention in each allocation round. The key design principle is that generative reasoning determines how operator intent influences service priorities, while recurring resource-allocation decisions remain independent of LLM inference. The main contributions of this paper are:

B. Edge Resource Orchestration and AI-Driven Control

Traditional optimization and heuristics. Resource orchestration for 5G edge network slicing has been extensively studied through centralized optimization frameworks and heuristic approaches. Mixed-Integer Linear Programming (MILP) has been extensively used to optimize Virtual Network Function (VNF) embedding and multi-domain resource allocation, aiming to minimize operational costs while satisfying capacity constraints [5], [6]. Concurrently, standardized Software-Defined Networking (SDN) and Network Function Virtualization (NFV) orchestrators, such as Open Source MANO (OSM) and ONAP, provide end-to-end slice lifecycle management across heterogeneous domains. At the cloud-native execution layer, Kubernetes-based platforms can be extended with Container Network Interface (CNI) mechanisms to provide tenant isolation and traffic shaping for network slices [7]. Despite their strengths, centralized orchestration approaches can face computational and responsiveness challenges as the optimization problem and number of managed resources grow. MILP-based formulations, for example, may incur substantial computational complexity for dynamic resource-allocation • A two-timescale orchestration architecture that separates LLM-based semantic intent translation from determinis- problems [5]. Conventional orchestration and scheduling mechatic MAS-based resource arbitration, keeping generative nisms typically operate on structured telemetry and predefined reasoning outside the recurring operational control loop; policies, providing limited support for directly interpreting high-level natural-language operator intents. This creates a • An intent-to-economic-policy mechanism that translates high-level operator intents into slice-specific priority, complementary challenge for edge orchestration: resource budget, income, and bidding-aggressiveness parameters decisions must remain responsive to changing infrastructure conditions while also reflecting service priorities expressed at consumed by autonomous slice agents; a higher semantic level. • An end-to-end implementation using SPADE, Prometheus, AI-driven and intent-based orchestration. AI-driven Kubernetes, Open5GS, and Calico, with resourceallocation decisions enforced directly on live UPF pods approaches have been widely explored to adapt networkslicing policies and resource allocation to changing conditions. through dynamic CPU and bandwidth updates; Deep Reinforcement Learning (DRL) and Multi-Agent RL • An experimental evaluation on a real 5G edge testbed across two complementary scenarios, each repeated 30 (MARL) can dynamically optimize resource distribution, but times, demonstrating intent-driven resource differentiation, training requirements, reward-design complexity, and variadaptation to changing demand, preservation of config- ability in learned policies can complicate predictable SLA ured minimum guarantees, and corresponding effects on enforcement [8]. More recently, researchers have investigated Large Language Models (LLMs) in Intent-Based Networking application-level QoE. (IBN) as semantic interfaces for translating natural-language This paper is organized as follows. Section II surveys related operator intents into structured network policies [9], [10]. approaches to network-slice orchestration and AI-driven control. LLM-based intent translation may also introduce additional Section III presents the proposed method. Section IV reports inference overhead compared with lighter models [10], which results and discussion. Section V concludes. becomes relevant for frequently executed control decisions. These characteristics motivate our two-timescale architecture: II. BACKGROUND & R ELATED W ORK an event-driven LLM Strategist translates operator intent into economic policy parameters, while recurring resource arbitraA. MAS tion is delegated to a deterministic MAS-based mechanism A Multi-Agent System (MAS) is a computational paradigm using current infrastructure telemetry. composed of autonomous, communicating software entities, reIII. P ROPOSED M ETHOD ferred to as agents, that perceive their environment, reason over local objectives, and coordinate with peers to solve problems As illustrated in Fig. 1, the challenge is not merely how to that benefit from distributed decision-making. Representing allocate resources for 5G network slices, but how to translate each network slice as an autonomous agent allows slice-level high-level operator intents into concrete infrastructure actions. state and resource requirements to be expressed independently, High-level intents are commonly realized through manual or while coordination mechanisms arbitrate contention over shared static configurations, while resource orchestration must arbitrate resources. dynamic contention over limited computing and bandwidth

resources. Our architecture addresses this gap through a two-timescale separation between semantic interpretation and operational resource arbitration. At the macro scale, a Large Language Model (LLM)-based Strategist interprets naturallanguage intents and translates them into economic policy parameters that express the relative priority of each slice. At the operational scale, autonomous Multi-Agent System (MAS) agents combine these parameters with current infrastructure telemetry to resolve resource contention through deterministic arbitration. In this way, operator intent influences service priorities without placing generative reasoning inside the recurring resource-allocation loop. A. Implementation and Use Cases

receiving a call-for-proposals (cfp) from the Resource agent, it maps its highest resource utilization into one of four stress tiers (high, moderate, comfortable, low) using configurable thresholds, and computes a priority-weighted coefficient that determines how aggressively it requests additional capacity. Priority p ∈ [1, 10] is normalized and passed through a cubic curve before being mapped into the tier’s coefficient range [cmin , cmax ] (as Eq. 1): 1 ϕ(p) = 2



p−5 5

3

! +1 ,

(1)

c(p) = cmin + ϕ(p) (cmax − cmin )

so that, under the same congestion level, high-priority slices 1) System Architecture: The proposed system comprises request larger capacity increments, while low-priority slices four cooperating agent roles, implemented with the SPADE remain closer to the lower bound of the coefficient range. The framework – specifically its SPADE-LLM extension for requested target for each resource is limit + base · c(p), and LLM integration – and communicating asynchronously over the bid itself is b = min(budget, β · m), where β is the slice’s XMPP using FIPA-inspired performatives (cfp, propose, current base bidding-aggressiveness parameter (already scaled accept-proposal, reject-proposal, inform). A by the Strategist-assigned weight from Section III-A2) and m Strategist agent uses the gpt-oss-20b LLM and interprets is a stress multiplier that further amplifies bidding under high natural-language network intents. A lightweight Dispatcher congestion and dampens it when the slice is comfortable. agent parses the Strategist’s structured output and routes The Resource agent collects all proposals within a fixed individual policy updates to the corresponding Slice agents, window and resolves contention using a Vickrey-inspired sealedone per network slice, each bound to a dedicated User bid, second-price mechanism [11]: the highest bidder is selected, Plane Function (UPF) pod. Periodically, a Resource agent while the corresponding price is determined by the secondacts as auctioneer: it runs a sealed-bid auction to resolve highest bid. The winner’s request is funded first from any free contention over the shared CPU and bandwidth pool, and capacity in the shared pool and, if insufficient, proportionally enforces the outcome directly on the underlying Kubernetes reallocated from the non-winning slices – never below each cluster. This separation confines the LLM’s comparatively slice’s configured minimum floor, with any residual shortfall slow and non-deterministic reasoning to event-driven policy redistributed starting from the least-resourced non-winning updates triggered by changes in operator intent, while recurring slices so that no active slice is driven below its guaranteed resource-arbitration decisions are executed deterministically by operating minimum. Algorithm 1 summarizes a single auction the MAS without further language-model inference. round. The outcome is enforced directly on the cluster: CPU 2) Intent-to-Policy Translation: The Strategist agent is requests and limits are updated via an in-place Kubernetes configured with a system prompt that describes the network’s pod resize, and bandwidth is enforced by patching the Calico slices, their default priority weights, and an explicit economic CNI QoS annotations on the affected UPF pods, closing the engine: for a target weight w ∈ [1, 10], the wallet capacity, loop from bid to infrastructure within the same auction period wealth-regeneration rate, and base bidding aggressiveness of without operator intervention. Memory utilization contributes a slice are respectively B0 · w, I0 · w, and β0 · w, where to the bidding decision but is not dynamically enforced in B0 , I0 , and β0 are operator-defined base constants. Given a the current prototype because updating memory limits in the natural-language intent (e.g., an operator reporting a publicpresent deployment would require restarting the target pod and safety emergency, as in Fig. 1), the LLM is constrained, via interrupting the network service. Dynamic memory enforcement structured output, to return one JSON object per affected slice is therefore left for future work. containing the target slice, its new priority weight, and a short justification. The Dispatcher agent validates and forwards each IV. R ESULTS AND D ISCUSSION update as an inform message to the corresponding Slice We evaluate the proposed architecture on a real Kuberagent, which immediately recomputes its budget, income, and netes/Open5GS edge cluster0 , whose UPF pods run on an base bidding aggressiveness from the new weight. Once these 8-core amd64 node with 8 GB of RAM and 35 GB of storage, parameters are updated, the operator intent is encoded into the running Ubuntu 24.04 LTS and Kubernetes v1.34.9, using economic leverage carried by the slice in subsequent auction two complementary experiment families: one exercises the rounds, which proceed independently of further LLM inference complete intent-to-infrastructure path, while the other isolates until a new intent is received. the operational resource-arbitration layer. Each experiment 3) Dynamic Bidding and Vickrey-Inspired Resource Arbitration: Each Slice agent continuously tracks its UPF’s CPU, 0 Full cluster specification available at https://computernetworks-ufrgs.githu memory, and bandwidth utilization via Prometheus. Upon b.io/k8s-cluster/.

Fig. 1. Problem Positioning.

Algorithm 1 Single-round auction and resource reallocation (firefighters > police > ambulance) does not follow either the Require: Slice agents S, shared pool free capacity, per-slice peak priority weight a slice ever received or its elevated-priority minimum floors duration in isolation. Firefighters combine the longest elevated 1: Resource agent broadcasts cfp to all s ∈ S window (420 s, t=180–600, ∼47% of the run) with the least2: for each s ∈ S (in parallel) do contested first two-thirds of it, before police also escalates at Measure utilization us from Prometheus; select stress t=420, and end with the highest average bitrate. Police hold 3: tier priority 9 for a shorter window (280 s, t=420–700, ∼31%), 4: Compute coefficient cs via Eq. (1) for the tier but benefit from firefighters stepping back down at t=600, 5: targets ← limits + bases · cs leaving the second half of their own window less contested. 6: bs ← min(budgets , βs · mtier ) The ambulance reaches the single highest priority weight of 7: Reply propose with (bs , targets , limits ) the experiment (10) but only from t=600 onward (∼33% of 8: end for the run) and, for part of that window, still competes against an 9: i∗ ← arg maxs∈S bs ▷ highest-bid slice elevated police (until t=700); it ends with the lowest average 10: price ← maxs̸=i∗ bs ▷ second-highest bid bitrate of the three despite its peak weight advantage. The 11: need ← targeti∗ − limiti∗ allocation outcome therefore depends jointly on a slice’s priority 12: Fund need from the free pool, then proportionally from non- history, its measured resource conditions, and the priorities of winning slices, respecting each non-winning slice’s min- simultaneously competing slices, rather than on priority alone. imum floor (remaining shortfall covered least-resourcedfirst) 13: Grant i∗ the allocated capacity; debit price from its budget Fig. 2 tracks bid, allocated bandwidth, and bitrate round-by14: Patch CPU (K8S resize) and bandwidth (Calico annota- round across the four intents (30 s per auction round). Each tions) on all affected UPF pods priority increase produces a bid and bandwidth response within one round, with bitrate lagging by roughly 2–4 rounds (60– 120 s); the corresponding bandwidth reduction falls disproporlasts 900 s and was repeated 30 times; the resource auction tionately on whichever slice is currently least elevated (event runs every 30 s (30 rounds per run), and all figures below are 2 reduces the ambulance’s allocated bandwidth by 51%, while averages across the 30 repetitions unless noted otherwise. The the firefighters’ allocation changes only marginally). Event 3 total CPU and bandwidth pools are fixed at 0.8 cores and coincides with the DASH content looping back to its lowest 35 Mbps, respectively, shared by three slices, each bound to a representation for all three slices; thus, the immediate bitrate response cannot be attributed solely to resource reallocation, DASH video client used as a QoE proxy. although the bandwidth traces show that the policy changes A. Scenario 1: Smart-City Intents had already propagated to the infrastructure. Event 4 produces In this scenario, three public-safety services (police, fire- a slower and more variable response than the preceding fighters, and an ambulance) compete for the shared pool, and priority increases. This indicates that the Strategist’s intent four natural-language intents are delivered to the Strategist shapes the allocation outcome without fully determining it: agent at fixed points in every run, with the corresponding the deterministic resource-arbitration mechanism continues to priority updates generated from the supplied intents rather than respond to each slice’s measured congestion, consistent with manually configured. the two-timescale design described in Section III. The summed Table I summarizes the aggregated QoE and resource- allocated bandwidth across the three slices averages 34.7 Mbps allocation statistics. The ranking observed across the full run of the 35 Mbps pool throughout the run, showing that the

TABLE I S MART-C ITY SCENARIO – AGGREGATED STATISTICS OVER 30 RUNS (900 S EACH ) Slice

Bitrate (kbps)

Buffer (s)

Stalls

Quality switches

Mean bid

BW (Mbps)

CPU (cores)

Ambulance Firefighters Police

5828 ± 3401 9748 ± 4381 6024 ± 3462

12.6 14.1 12.7

0.45 0.57 0.56

17.5 13.0 19.5

62.3 59.7 62.0

9.8 15.0 9.9

0.067 0.105 0.067

Ambulance

1

Firefighters

2

3

Police

Bronze

4

1

BW (Mbps)

Bid value

150 100 50

4

5

6

30 20 10

15000

Bitrate (kbps)

25

Allocated BW (Mbps)

3

Silver

0

0

20 15 10

10000 5000 0

5

0

0 15000

Bitrate (kbps)

2

Gold

100 200 300 400 500 600 700 800 900

Time (s)

Fig. 3. Allocated bandwidth and bitrate per slice (mean of 30 runs; shaded band is ±1 std. dev.). Events: (1) Gold paused, t≈121 s; (2) Silver paused, t≈272 s; (3) Gold resumed, t≈422 s; (4) Bronze paused, t≈541 s; (5) Silver resumed, t≈662 s; (6) Bronze resumed, t≈782 s.

10000 5000 0 0

100 200 300 400 500 600 700 800 900

Time (s)

Fig. 2. Bid, allocated bandwidth, and bitrate per slice over the four Smart-City intents (mean of 30 runs; shaded band is ±1 std. dev.). Events: (1) fire report, t=180 s, firefighters → 9; (2) police chase, t=420 s, police → 9; (3) medical emergency and end of fire, t=600 s, ambulance → 10, firefighters → 5; (4) stand-down, t=700 s, police → 5.

mechanism allocates nearly the entire configured bandwidth capacity without exceeding the shared-pool limit. B. Scenario 2: Hierarchical Slices This scenario isolates the operational resource-arbitration mechanism from the Strategist layer: three slices are assigned static contractual priorities – Gold (8.0), Silver (5.0), and Bronze (3.0) – and traffic is paused and resumed on each slice in turn to emulate subscriber disconnection and reconnection, without invoking the Strategist.

Table II shows clear service differentiation across the three priority classes. Gold’s average bitrate is 3.09× Bronze’s, exceeding the nominal priority ratio of 8.0/3.0 ≈ 2.67×. This amplification is consistent with the cubic priority-to-coefficient mapping ϕ(p) in Eq. (1), which intentionally strengthens differentiation toward the upper end of the priority range. QoE indicators follow the same ordering: Gold records the lowest stall rate (0.09) and fewest quality switches (10.6), compared with 0.65 stalls and 19.7 switches for Bronze. Fig. 3 shows the reallocation mechanism directly. Following event 1, Gold’s bandwidth falls from about 21.0 to 7.6 Mbps over 300 s (ten rounds) as the Resource agent redirects freed capacity to Silver and Bronze; after event 3, it takes about six rounds (∼180 s) to climb back above 20 Mbps, tracking the ramp-up of its own measured congestion rather than an instantaneous restoration. Bronze’s pause (event 4) drives its allocation toward the configured 1.0 Mbps floor, and it remains close to that floor even after resuming (event 6), as Gold and Silver’s combined demand continues to dominate the pool. Summed bandwidth remains close to the 35 Mbps pool throughout, including while traffic is simultaneously paused

TABLE II H IERARCHICAL SLICING SCENARIO – AGGREGATED STATISTICS OVER 30 RUNS (900 S EACH ) Slice

Bitrate (kbps)

Buffer (s)

Stalls

Quality switches

Mean bid

BW (Mbps)

CPU (cores)

Bronze Silver Gold

3673 ± 2885 4149 ± 2936 11347 ± 4477

11.6 12.1 16.0

0.65 0.47 0.09

19.7 12.9 10.6

25.6 32.1 52.8

5.3 9.7 19.7

0.030 0.026 0.088

on two slices, showing that configured minimum floors are preserved while the available bandwidth pool is dynamically reassigned as slice demand changes.

while the Hierarchical experiments demonstrated prioritydependent resource and QoE differentiation, adaptation to changing demand, and preservation of configured minimum guarantees. C. Discussion The mechanism’s economic parameters were empirically Taken together, the two scenarios evaluate complementary configured for the evaluated scenarios, and their generalization aspects of the proposed two-timescale architecture. The Smart- across different slice mixes and traffic patterns remains to be City scenario exercises the full intent-to-infrastructure path, investigated. Future work includes enabling slice agents to showing that natural-language operator intents translate into learn these parameters through reinforcement learning without economic policy parameters and are subsequently reflected interfering with the deterministic resource-arbitration loop [8], in slice bids and bandwidth allocations. The Hierarchical and allowing the Strategist to incorporate utilization telemetry scenario isolates the operational resource-arbitration layer, to reconcile operator intents proactively with observed network showing priority-dependent resource and QoE differentiation conditions. under changing demand. In both scenarios, the final allocation ACKNOWLEDGMENT remains responsive to measured resource conditions rather than being directly prescribed by the LLM, illustrating the intended The authors thank FAPESP MCTIC/CGI Research projects separation between semantic policy formation and recurring 2018/23097-3 - SFI2: Slicing Future Internet Infrastructures and infrastructure control. 20/05182-3 - PORVIR-5G: programability, orchestration and The CPU figures reported in Tables I and II (0.03–0.11 cores) virtualization in 5G networks, and Fundação para a Ciência reflect the CPU demand observed for the slice workloads e a Tecnologia (FCT) within the R&D Unit Project Scope during the experiments. The DASH traffic used in both UID/00319/2025 - Centro ALGORITMI (ALGORITMI/UM) scenarios is primarily bandwidth-driven, which explains the for supporting this work. comparatively low CPU demand. These figures do not include R EFERENCES the computational cost of the Strategist’s LLM inference, which was not instrumented in this deployment. Because the Strategist [1] J. S. B. Martins, T. C. Carvalho, R. Moreira, C. B. Both, A. Donatti, J. H. is invoked only when a new intent arrives rather than in every Corrêa, J. A. Suruagy, S. L. Corrêa, A. J. G. Abelem, M. R. N. Ribeiro, J.-m. S. Nogueira, L. C. S. Magalhães, J. Wickboldt, T. C. Ferreto, allocation round, this cost is intermittent and remains outside R. Mello, R. Pasquini, M. Schwarz, L. N. Sampaio, D. F. Macedo, J. F. the recurring resource-arbitration loop. We do not claim the De Rezende, K. V. Cardoso, and F. De Oliveira Silva, “Enhancing network overall architecture has lower computational overhead than a slicing architectures with machine learning, security, sustainability and experimental networks integration,” IEEE Access, vol. 11, pp. 69 144– static baseline; quantifying the LLM’s inference overhead, and 69 163, 2023. weighing it against the intent-aware differentiation it enables, [2] R. Moreira, L. F. Rodrigues Moreira, T. C. Carvalho, and is left for future work. F. de Oliveira Silva, “Noisy neighbor influence in the data plane of V. C ONCLUDING R EMARKS This paper presented a two-timescale architecture for autonomous resource orchestration in 5G edge network slicing. An LLM-based Strategist provides a semantic translation layer between natural-language operator intent and slicespecific economic policy parameters, while a deterministic Multi-Agent System performs recurring resource arbitration through a Vickrey-inspired second-price mechanism. This separation combines semantic flexibility at the policy layer with deterministic execution at the operational resource-arbitration layer, without requiring generative inference for each allocation decision. Evaluated end-to-end on a real Kubernetes/Open5GS edge cluster over 30 repetitions per scenario, the SmartCity experiments showed operator intents propagating into measurable changes in slice bids and bandwidth allocations,

beyond 5g networks,” in 2026 IEEE 23rd Consumer Communications & Networking Conference (CCNC), 2026, pp. 1–6. [3] R. Moreira, R. Pasquini, J. S. Martins, T. C. Carvalho, and F. de Oliveira Silva, “Ai-driven orchestration at scale: Estimating service metrics on national-wide testbeds,” Future Generation Computer Systems, vol. 174, p. 107971, 2026. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0167739X25002663 [4] M. Andrade and J. Wickboldt, “A study on 5g network slice isolation based on native cloud and edge computing tools,” Journal of Network and Systems Management, vol. 33, no. 4, p. 90, Jul 2025. [Online]. Available: https://doi.org/10.1007/s10922-025-09958-5 [5] N. Ejaz and S. Choudhury, “A comprehensive survey of linear, integer, and mixed-integer programming approaches for optimizing resource allocation in 5g and beyond networks,” arXiv preprint arXiv:2502.15585, 2025. [6] A. Jalalian, S. Yousefi, and T. Kunz, “Network slicing in virtualized 5g core with vnf sharing,” Journal of Network and Computer Applications, vol. 215, p. 103631, 2023. [7] A. T. de Cock Buning, I. Vidal, and F. Valera, “Architecting multi-cluster layer-2 connectivity for cloud-native network slicing,” Future Internet, vol. 18, no. 1, p. 39, 2026.

[8] Z. Cui, F. Qamar, S. H. A. Kazmi, K. A. Zainol Ariffin, G. A. Safdar, and M. H. ur Rehman, “A review of multi-agent deep reinforcement learning for resource allocation in beyond 5g network slicing: solutions, challenges and future research directions,” PeerJ Computer Science, vol. 12, p. e3728, 2026. [9] A. Mekrache, A. Ksentini, and C. Verikoukis, “Intent-based management of next-generation networks: an llm-centric approach,” IEEE Network, vol. 38, no. 5, pp. 29–36, 2024. [10] W. L. Phone, B. El Boudani, T. Dagiuklas, and S. Ghosh, “Performance comparison of ibn orchestration using llm and slms,” arXiv preprint arXiv:2603.06647, 2026. [11] W. Vickrey, “Counterspeculation, auctions, and competitive sealed tenders,” The Journal of Finance, vol. 16, no. 1, pp. 8–37, 1961.

Related documents

• Conflict-Aware Autonomous Resource Orchestration for 5G Edge Slicing Using Multi-Agent Systems and LLMs#1325799
• METIS: A Declarative Slice Orchestrator for Application-Centric 5G/6G Networks#422213
• Advanced LLM-Enhanced Intent-Based 5G Network Management using Dynamic Semantic Routes#482098
• AI-Native Orchestration in the 6G Continuum: Evolving Operator Platforms with Agentic AI#667937
• Enabling Parameter-Efficient Multi-Tasking via Multi-Head Architecture for Edge Device Deployment#168976