Agoran SRB: 6G RAN Automation Marketplace

• Agoran SRB is defined as an agentic, stakeholder-centric marketplace for 6G RAN automation that transforms natural-language intents into deployable resource policies.
• The system employs a tripartite AI governance model—Legislative, Executive, and Judicial—to ensure regulatory compliance, real-time observability, and trust-based arbitration.
• Experimental deployment on a 5G testbed demonstrated a 37% increase in eMBB throughput, 73% reduction in URLLC latency, and an 8.3% saving in PRB usage compared to static baselines.

Searching arXiv for the Agoran SRB paper and the cloud bartering paper for citation support. Agoran Service and Resource Broker (SRB) is an agentic, open marketplace for 6G Radio Access Network (RAN) automation and resource brokering, inspired by the ancient Greek agora and designed to bring stakeholders directly into the operational loop. It is presented as a response to the limitation that today’s network slice controllers remain rigid, policy-bound, and unaware of the business context. In Agoran, authority is distributed across three autonomous AI branches, while stakeholder-side Broker Agents (bAgents) negotiate service and resource allocations from natural-language intents to a consensus intent that is deployed to Open and AI RAN controllers. On a private 5G testbed evaluated with realistic traces of vehicle mobility, the reported system achieved a 37% increase in throughput of eMBB slices, a 73% reduction in latency of URLLC slices, and an end-to-end 8.3% saving in PRB usage compared to a static baseline (Chatzistefanidis et al., 5 Aug 2025).

1. Conceptual Scope and Operational Model

Agoran is defined as an agentic marketplace for 6G RAN automation in which business entities such as Mobile Network Operators (MNOs), network slice customers, and vertical service providers actively participate in real-time service and resource allocation via autonomous agents. Its design is explicitly stakeholder-centric and ultra-flexible, with intent-based multi-domain operation: stakeholders express high-level intents in natural language, these are translated into structured, machine-readable offers, and the offers are resolved collectively into deployable policy. The central architectural image is a “central agora” in which customer bAgents interact while legislative, executive, and judicial branches perform compliance, arbitration, and enforcement roles (Chatzistefanidis et al., 5 Aug 2025).

The operational significance of this framing is that service brokering is no longer treated as a static mapping from SLA templates to controller commands. Instead, planning, deployment, and optimization are all embedded in an autonomous deliberative loop. A common misunderstanding is to read the SRB as merely a policy gateway; the reported design instead keeps stakeholders in the loop throughout the lifecycle of service negotiation and adaptation. This suggests a shift from brokered admission control toward brokered collective governance.

2. Tripartite Governance Architecture

The most distinctive feature of Agoran is its tripartite governance model, in which authority and decision-making are split among three mutually independent, AI-driven power branches, mirroring societal separation of powers. Each branch is associated with a specific class of tasks and data dependencies, and each contributes to a distinct control function within the marketplace (Chatzistefanidis et al., 5 Aug 2025).

Branch	Core role	LLM agent tasks
Legislative	Defines “network law”: contracts, standards, etc.	Uses RAG LLMs to answer compliance queries and enforce regulatory alignment
Executive	Monitors KPIs, enforces policy	Maintains real-time context via watcher-driven vector store; analyzes telemetry for real-time policy updates
Judicial	Assesses trust, arbitrates, incentivizes	LLM-based content moderation filters, trust score calculation, applies incentives/penalties to agent behaviors

The Legislative branch, implemented through an lAgent, maintains an evolving corpus of spectrum regulations, policies, and contractual clauses. Upon a compliance query such as whether an offer is legal, it retrieves relevant regulatory snippets, injects them into an LLM prompt, and returns a citation-rich answer; the knowledge base updates with new precedents. The Executive branch, through an eAgent, provides what the paper terms “agentic observability” by maintaining situational awareness in a watcher-updated vector database. Kubernetes watchers and telemetry agents push updated telemetry, configurations, and counters into a vector store, and agents query either the vector store or an external monitoring API to ground decisions in current network state. The Judicial branch, through a jAgent, embeds every message and draft intent, classifies content as benign, borderline, or malicious by referencing a library of manipulation and toxicity patterns, and issues warnings or temporary penalties accordingly.

Taken together, these branches decouple regulatory interpretation, runtime observability, and trust arbitration. This separation is important because it prevents unilateral control by any single decision module while preserving automated operation.

3. Negotiation Pipeline and Consensus Formation

The negotiation process begins with expression: stakeholders submit natural-language intents to their bAgents through a human-to-broker interface. Offer generation follows, in which bAgents structure the intents into KPI-bound proposals, while the SRB-side Mediator Agent runs a multi-objective evolutionary optimizer to generate a set of feasible, Pareto-optimal candidate offers. Agents then debate and select among these offers, justifying their preferences, and a consensus is reached typically in a single round. The consensus is then decomposed into actionable policies and enforced by RAN controllers through standardized interfaces such as O-RAN A1, E2, and SBI; the system is also described as vendor-neutral through JSON APIs (Chatzistefanidis et al., 5 Aug 2025).

Several properties of this workflow are emphasized. First, the process is fully machine-driven after the initial intent input. Second, one-round consensus is presented as a signature of process efficiency rather than as an approximation achieved by repeated back-and-forth bargaining. Third, standards alignment is treated as a deployment property rather than a conceptual aspiration: the system is described as operationally compatible with current and future O-RAN and AI-RAN deployments.

The negotiation structure therefore couples language-based stakeholder input with formally generated solution sets. This suggests that natural-language intent is not used as a replacement for optimization, but rather as a front end for constrained multi-objective search and mediated agreement.

4. Optimization, Trust Score, and Incentive Logic

The core optimization engine is NSGA-II based and is used to generate a set of non-dominated, Pareto-optimal SLA offers balancing multi-stakeholder objectives and technical constraints. The resource vector per slice is given as $\mathbf{x}_i = (b_i, c_i, p_i, s_i)$ for bandwidth, compute, power, and storage, and the joint decision vector is $$\mathbf{x} = (\mathbf{x}_e, \mathbf{x}_u, \mathbf{x}_m)$$ for eMBB, URLLC, and mMTC slices. Global resource limits and per-slice minimum and maximum KPIs constrain the search space. The KPI models are reported as follows (Chatzistefanidis et al., 5 Aug 2025):

$T_i(b_i, m_i) = \kappa Q_{m_i} R_{m_i} b_i$

$$L_i(b_i, m_i) = L_i^{fix} + \frac{1}{\mu_i(b_i, m_i)(1-\rho_i)}$$

$C_i = \alpha (c_i + s_i)$

$E_i = p_i$

The optimization objective is

$$\mathbf{f}(\mathbf{x}) = \left( -\sum_i T_i, \sum_i L_i, \sum_i C_i, \sum_i E_i \right),$$

so the optimizer maximizes throughput while minimizing latency, cost, and energy. Each candidate is a 12-dimensional vector, and evolutionary search with crossover, mutation, and constraint repair returns a Pareto front $\mathcal{P}^*$; the top-$k$ offers from this front populate the structured JSON negotiation space.

Agoran’s trust framework is built around a composite Trust Score $T$ that combines Satisfaction and Coherence:

$$\mathbf{x} = (\mathbf{x}_e, \mathbf{x}_u, \mathbf{x}_m)$$0

with Satisfaction defined by weighted penalties for deviation from valid offers, deviation from intent, and deviation from mediator recommendation:

$$\mathbf{x} = (\mathbf{x}_e, \mathbf{x}_u, \mathbf{x}_m)$$1

Coherence aggregates factual accuracy, logical consistency, and semantic coherence of the agent’s rationale, with the highest weight placed on factual accuracy. The incentive and disincentive engine then maps outputs into operational consequences: benign output triggers no action, borderline output triggers a warning, malicious or hallucinatory output triggers a fine or penalty reducing negotiation influence, and consistently constructive behavior yields credits for future bargaining (Chatzistefanidis et al., 5 Aug 2025).

This combination of Pareto-front generation and judicial arbitration is central to the SRB’s design. The optimizer constrains the feasible space of offers, while the judicial mechanism constrains the discourse used to choose among them.

5. Experimental Deployment and Reported Results

The reported evaluation was conducted on a real OpenAirInterface 5G NR cell with 40 MHz bandwidth and peak 133.7 Mbps, using FlexRIC control and containerized microservices for SRB components; LLMs ran locally and in the cloud. The workload consisted of eMBB slices optimized for high throughput, URLLC slices for low latency, and mMTC slices for low cost and energy. The scenario included multiple phases with dynamic channel conditions based on realistic MCS traces, intent swaps, and sudden RAN shut-offs, with the purpose of testing robustness and adaptability under real-world-like conditions (Chatzistefanidis et al., 5 Aug 2025).

Metric	Reported outcome
eMBB throughput	37% increase
URLLC latency	73% reduction
End-to-end PRB usage	8.3% saving
PRB dynamics	24% reduction in low demand, 16% increase in high demand
1B model efficiency	approximately 80% of GPT-4.1’s decision quality, 6 GiB memory, 1.3 seconds

The negotiation and model-quality results are also specific. A fine-tuned 8B LLaMA model achieved about 86% of GPT-4.1 in multi-agent negotiation, while the 1B model reached approximately 80% within 1.3 seconds and 6 GiB VRAM after being fine-tuned for five minutes on 100 GPT-4 dialogues. Trust Score analysis indicated that only large 8B+ models achieved high trustworthiness, whereas smaller models below 3B faltered in factual and logical coherence. The one-round consensus property was observed for all agents, excluding initial intent submission.

The three governance branches were also individually benchmarked. Legislative RAG achieved 100% retrieval accuracy for compliance queries in sub-2s on commodity hardware. Judicial arbitration detected toxic or malicious dialogue with F1 up to 0.96 and imposed incentives without significant latency overhead, reported as less than 0.4s. Executive observability achieved telemetry synchronization and agentic query answering in under 100 ms, supporting use in near-real-time RIC loops (Chatzistefanidis et al., 5 Aug 2025).

6. Relation to Brokering Traditions and Interpretive Significance

The term “broker” has prior usage in cloud and inter-cloud resource management, but the technical meaning differs substantially across systems. In the cloud-resource bartering literature, traditional brokers, including “Cloud Federation” and “Agoran Service Broker,” are characterized as relying on monetary pricing, often using FCFS or simple price-based allocation, with simple matching and minimal negotiation; in contrast, the Cloud Resource Bartering System (CRBS) introduces barter credits, urgency-aware allocation, and multi-agent negotiation without direct money (Gohera et al., 2018). Agoran SRB, by comparison, is framed not around barter credits or FCFS allocation, but around stakeholder-side negotiation agents, a Mediator Agent, Pareto-optimal offer generation, tripartite AI governance, and deployment into Open and AI RAN controllers (Chatzistefanidis et al., 5 Aug 2025).

This distinction matters because it prevents an imprecise identification of Agoran with earlier service-broker abstractions. Agoran is not described as a conventional intermediary for static matching or purely monetary exchange. It is instead a governance-capable marketplace in which compliance, observability, optimization, arbitration, and incentives are co-designed. A plausible implication is that “broker” now denotes a broader coordination substrate for intent resolution under technical, business, and regulatory constraints.

The broader significance claimed for Agoran is that it offers a concrete, standards-aligned path toward ultra-flexible, stakeholder-centric 6G networks. That significance rests on a specific combination of properties reported in the paper: continuous stakeholder involvement, one-round consensus, formal multi-objective optimization, real-time observability, trust-score-based arbitration, and measurable gains in throughput, latency, and PRB efficiency on a live testbed (Chatzistefanidis et al., 5 Aug 2025).