Agentic AI-Native 6G Architecture

• Agentic AI-Native 6G is a network paradigm that integrates autonomous, intent-driven agents with semantic and reasoning-native control across all network layers.
• It leverages a layered architecture combining deterministic infrastructure with added semantic abstraction and hierarchical multi-agent coordination for closed-loop adaptation.
• Empirical evaluations demonstrate improved throughput, energy efficiency, and operational accuracy in domains such as O-RAN, SAGIN, and federated learning.

Agentic AI-Native 6G denotes a family of 6G architectures in which intelligence is treated as a native network capability and realized through autonomous, goal-driven agents rather than isolated optimization modules or static rule engines. Across recent formulations, this paradigm combines intent-based control, semantic abstraction, distributed multi-agent coordination, tool-mediated execution, and closed-loop adaptation over device, edge, RAN, and core domains. In this view, 6G is no longer only a substrate for transporting bits or semantic features; it becomes a control fabric that can interpret natural-language intents, reason over multimodal state, execute network actions through standardized tools and interfaces, and continuously refine behavior through feedback, memory, simulation, or digital twins (Li et al., 19 Feb 2026, Seo et al., 19 Feb 2026, Ferrag et al., 2 May 2026).

1. Conceptual foundations

The most compact definition treats agentic AI in 6G as a shift from offline, task-specific, goal-static machine learning toward autonomous, interactive, closed-loop systems embedded directly in the 6G stack. A canonical agent is organized around a perception–decision–action loop: perception gathers multimodal observations such as CSI, KPIs, and textual intents; decision reasons over current state, historical memory, and goals, possibly invoking tools; action configures network or physical-layer modules; and feedback returns performance metrics and environment reactions for policy updates (Li et al., 19 Feb 2026). This formulation differs from conventional AI-for-networks, where learning models are typically attached to narrow tasks and execute static mappings learned offline.

A second conceptual strand reframes communication itself. In reasoning-native agentic communication, the operative unit is no longer only bits or semantic features, but cognitive state or intentional tokens, and the system objective becomes collaborative decision coherence rather than transmission fidelity alone. The central difficulty is belief divergence: two autonomous agents may decode the same message correctly yet act inconsistently because their internal beliefs and policy updates evolve differently. The proposed response is a coordination plane that triggers communication when predicted misalignment in internal belief states exceeds a tolerable bound, rather than when channel conditions or raw data changes alone warrant transmission (Seo et al., 19 Feb 2026).

A third strand emphasizes semantic-native and intent-native operation. In this formulation, communication is organized around task-relevant semantics, semantic fidelity, and task success rate, while network entities become agents with goals, autonomy, and coordination capabilities distributed across PHY/MAC, near-real-time RIC, and non-real-time RIC. This suggests that agentic AI-native 6G is not a single architecture but a convergence of three closely related ideas: intent-aware control, semantic abstraction, and reasoning-centric coordination (Feng et al., 4 Dec 2025). A broader systems argument makes the same point from the opposite direction: 6G cannot rely only on optimization-centric control loops because heterogeneous services, policy constraints, trust requirements, and cross-domain interactions require bounded reasoning entities operating above deterministic 3GPP infrastructure (Ferrag et al., 2 May 2026).

2. Architectural patterns and control planes

A recurrent architectural pattern is a layered stack in which deterministic infrastructure is preserved, while semantic and agentic control layers are added above it. One four-layer formulation separates deterministic network infrastructure, semantic abstraction of intent and context, hierarchical reasoning, and a distributed multi-agent fabric spanning device, edge, and core domains. In that design, agents are policy-governed reasoning entities, and execution remains tool-mediated and bounded by deterministic infrastructure interfaces (Ferrag et al., 2 May 2026).

O-RAN provides a particularly explicit realization of this multi-timescale view. A multi-scale framework places an LLM agent in the Non-RT RIC for intent translation, policy creation, and model lifecycle management; SLM agents in the Near-RT RIC for low-latency optimization and orchestration of xApps; and WPFM agents near the distributed unit for fast inference close to the air interface. These agents cooperate through A1, E2, O1, a Model Catalog, a Data Lake, a Knowledge Base, Analytics, and a Simulation/Digital Twin environment, thereby turning O-RAN’s existing open interfaces into tools for a hierarchical agentic control system (Navidan et al., 15 Feb 2026).

The same architectural principle appears in mobile core networks, but with different interface choices. One prototype frames “tool use as action,” with MCP as the agent-to-tool interface and A2A as the inter-agent protocol. In that design, host, monitoring, and execution agents coordinate over A2A, discover tools through tools/list, invoke them through tools/call, and let those tools translate agent actions into SBI calls or system-level operations on network functions such as the AMF and NRF (Garigipati et al., 4 May 2026). A related core-network proposal extends the 3GPP architecture with an Agentic AI-Native layer comprising a Network Planner Agent, a Network Executor Agent, an Intent Manager, monitor tools, execution tools, and an MCP server, thereby establishing a closed loop from high-level intents to CP/DP and MANO actions over existing xGC APIs (Barbosa et al., 29 May 2026).

Slice orchestration architectures generalize the same control-plane logic across economic and operational dimensions. One layered control plane organizes a Trading Agent, a Policy and Safety Agent, a Slice Orchestration Planning Agent, a Slice Orchestration Agent, and an SLA Monitoring Agent above Kubernetes-native infrastructure and 5G core and RAN resources. A natural-language interface implemented using MCP, together with a multi-model consortium of fine-tuned LLMs governed by a dedicated reasoning model, supports slice planning, deployment, monitoring, and economically informed decision-making in a single control function (Bandara et al., 27 Jan 2026).

3. Intent, semantics, and multimodal grounding

Intent is the principal abstraction through which agentic AI-native 6G connects human goals to executable network behavior. In one hierarchical multi-agent IBN framework, the network state is represented through RAN sectors, core nodes, and a latency matrix, while the final slice configuration is a tuple $\mathcal{A}^* = (r^*, b^*, c^*)$ selecting a sector, a band, and a UPF node. Natural-language intent is translated into QoS class, constraints, and weights in an engineering utility function, after which an orchestrator agent coordinates RAN and Core specialist agents through ReAct-style reasoning to synthesize feasible slice configurations (Jiang et al., 10 Jan 2026). This is an explicit example of an AI-native semantic layer: the system does not receive a complete optimization problem; it infers one from text.

At the physical layer, the same principle appears in a more tightly grounded form. AgenCom treats textual intent as a label for a point in a multi-dimensional objective space, jointly conditioned on CSI. Its multimodal perception module fuses CSI and user intent into a shared embedding; a GPT-2 Medium backbone with 355M parameters and a domain adapter maps that representation to a structured sequence of sub-actions selecting coding scheme, coding rate, modulation order, power compensation, precoding method, and channel estimation and equalization techniques; and a Sionna-based tool executor validates the chosen strategy and returns BER, achievable rate, and $p_{\text{extra}}$ (Li et al., 19 Feb 2026). In this formulation, the LLM is not an unconstrained language interface but the reasoning core inside a domain-adaptive policy network.

Semantic-native AI-RAN extends the same grounding problem across the RAN stack. A unified taxonomy organizes recent work along semantic abstraction level, agent autonomy and coordination granularity, and control placement. Semantic representations range from bit/symbol to feature/latent, intent/task, and knowledge/graph; agents range from non-agentic and single-agent to multi-agent and hierarchical; and control spans PHY/MAC local loops, near-RT RIC, non-RT RIC, and cross-layer end-to-end coordination. This suggests that intent-awareness in 6G is inseparable from the choice of semantic representation and the placement of control intelligence (Feng et al., 4 Dec 2025).

A reasoning-native view adds an additional distinction: semantic agreement is not sufficient if internal beliefs still diverge. In that perspective, communication is activated according to predicted misalignment in internal belief states, and a shared ontology becomes the structural anchor for both semantic tokenization and bounded recursive belief modeling. A plausible implication is that future agentic AI-native 6G systems may need to combine semantic abstraction with explicit models of belief and policy alignment rather than treating them as interchangeable (Seo et al., 19 Feb 2026).

4. Representative realization pathways and domains

The physical layer is one realization pathway, but not the only one. AgenCom shows how a single architecture can implement coherent intent-conditioned trade-offs across BER, achievable rate, and power compensation under heterogeneous Sionna-RT channels and three intent classes—high throughput, high reliability, and energy-aware—without separate hand-crafted policies (Li et al., 19 Feb 2026). This makes the physical layer a concrete test case for the broader claim that natural-language intent can be grounded into low-level PHY actions.

Space–air–ground integrated networks provide another realization path. An agentic SAGIN framework embeds LLM-based agents into a MAPE-K control plane with semantic resource perceivers, intent-driven orchestrators, and adaptive learners. The key mechanism is hierarchical agent–RL collaboration, in which the orchestrator dynamically shapes reward functions for RL agents according to semantic network conditions and operator intents. In a UAV-assisted AIGC orchestration scenario with 3 LEO satellites, 5 UAVs, 2 ground BSs, and 50 concurrent AIGC tasks, LLM-driven reward shaping produced 14% lower normalized UAV energy consumption than fixed-reward D3PG while also achieving the lowest average service latency (Zhang et al., 17 Mar 2026).

Federated learning over 6G yields a control-plane-oriented realization. There, Agentic AI is organized as retrieval, planning, coding, and evaluation agents that jointly manage client selection, incentive structuring, scheduling, resource allocation, adaptive local training, and code generation. The underlying premise is that FL over 6G is not only a learning problem but a combined task of learning and network management, so the agentic layer continuously refines both learning logic and network control using monitoring tools, optimization methods, closed-loop evaluation, and memory (Nguyen et al., 10 Mar 2026).

Agentic AI-RAN for low-altitude wireless networks offers a task-oriented SC3 realization. In that design, an edge node becomes an “agentic brain” for sensing, communication, computing, and control, using a Cognitive Planner with Chain-of-Thought reasoning, Contextual Memory, a Toolbox abstraction, MIG partitioning, and containerized deployment on a general-purpose GPU platform. In the demonstrated autonomous drone navigation task, the system achieved stable indoor navigation with closed-loop delay in the 500–680 ms range, while MIG partitioning separated communication and multimodal inference and enabled robust bidirectional communication under dynamic runtime conditions (Sun et al., 23 Jan 2026).

A more expansive networking vision is given by AgentNet, which treats the network itself as an ecosystem supporting interaction, collaborative learning, and knowledge transfer among foundation model-as-agents, embodied model-as-agents, and hybrid model-as-agents. In this view, generative foundation model-as-agent components act as interactive knowledge bases that can bootstrap embodied agents through synthetic data generation and environment simulation in digital-twin-based industrial automation and metaverse-based infotainment scenarios (Xiao et al., 20 Mar 2025). This suggests a broader reading of agentic AI-native 6G in which the network is designed not only to optimize itself, but to support a population of heterogeneous autonomous agents.

5. Empirical evidence and evaluation regimes

Evaluation in this area is heterogeneous because the literature measures different aspects of autonomy: behavioral coherence, operational correctness, PHY trade-offs, self-optimization gains, and management usability. Even so, a common pattern is visible: agentic methods are usually assessed by end-to-end task effects rather than solely by raw communication metrics.

A reasoning-native communication study makes that shift explicit through reasoning-centric KPIs. In collaborative humanoid manipulation, the reported Reasoning Alignment Score was 42% for classical communication, 68% for semantic communication, and 94% for agentic mutual-agentic-reasoning communication. In the same setting, Decision Impact per Bit was 3.5× higher than semantic, Mutual Belief Stability over 1000 cycles remained above 91% for the agentic scheme while semantic dropped to about 60%, and signaling overhead was 58% relative to a semantic baseline normalized to 100% (Seo et al., 19 Feb 2026). These results directly operationalize the claim that communication should regulate distributed reasoning rather than merely deliver representations.

Operational management prototypes evaluate a different axis: whether agentic systems can match human operators in practical workflows. On a live 5G Open RAN testbed, MX-AI was evaluated on 50 realistic operational queries and attained a mean answer quality of 4.1/5.0, 100% decision-action accuracy, and 8.8 seconds end-to-end latency with GPT-4.1 (Chatzistefanidis et al., 8 Aug 2025). In a separate AI-RAN co-management assistant, three capability layers were reported with 78% accuracy for design and planning a service, 89% accuracy for operating specific AI-RAN tools, and 67% accuracy for tuning AI-RAN performance, at an average response time of 13 seconds; the same study also reported a 43% hallucination rate, making clear that usability gains coexist with unresolved reliability issues (Srinivasan et al., 14 Feb 2026).

Reflection-driven self-optimization provides evidence for simulation-in-the-loop autonomy. A four-agent framework with scenario, solver, simulation, and reflector agents reported 17.1% higher throughput in interference optimization, 67% improved user QoS satisfaction through intent recognition, and 25% reduced resource utilization during low-traffic periods while maintaining service quality (Hu et al., 8 Dec 2025). In O-RAN slice control, a multi-scale LLM–SLM hierarchy increased VIP slice throughput by about 6% over an SLM-only controller while maintaining low average latency of about 22 ms for a latency-sensitive slice (Navidan et al., 15 Feb 2026).

The following examples summarize selected reported outcomes.

System	Setting	Reported result
Reasoning-native agentic communication	Collaborative humanoid manipulation	RAS 94%, DIB 3.5×, MBS >91%, overhead 58% (Seo et al., 19 Feb 2026)
MX-AI	50 realistic operational queries on live Open RAN	4.1/5.0 answer quality, 100% decision-action accuracy, 8.8 s E2E (Chatzistefanidis et al., 8 Aug 2025)
Reflection-driven self-optimization	RAN simulation-in-the-loop	17.1% higher throughput, 67% improved QoS satisfaction, 25% reduced resource utilization (Hu et al., 8 Dec 2025)
SAGIN agentic orchestration	UAV-assisted AIGC services	14% lower normalized UAV energy consumption and lowest average service latency (Zhang et al., 17 Mar 2026)

These results do not imply a single dominant architecture. Rather, they show that “agentic” is being validated through multiple, domain-specific criteria: PHY execution quality, cognitive alignment, operational correctness, self-optimization, and energy–latency trade-offs. This suggests that any unified evaluation framework for agentic AI-native 6G will need to span both network-centric and reasoning-centric metrics.

6. Challenges, controversies, and research agenda

Several limitations recur across the literature. One is domain mismatch: generic LLMs do not natively encode channel models, signal processing constraints, protocols, or safe action spaces. Physical-layer work therefore emphasizes domain adapters, structured action modeling, simulators, and curriculum or staged training, while reasoning-native communication emphasizes shared ontologies and bounded recursive belief modeling to prevent uncontrolled reasoning depth (Li et al., 19 Feb 2026, Seo et al., 19 Feb 2026). A second is the fundamental tradeoff between reasoning capability and system efficiency. An empirical study of 20 models on 6G-Bench reported that no single model simultaneously satisfied latency, throughput, and accuracy requirements, and that quantization had non-uniform effects across models, implying that heterogeneous deployment across device, edge, and core is necessary (Ferrag et al., 2 May 2026).

Safety and trust remain central controversies. Tool-grounded core-network prototypes showed that protocol overhead through MCP, A2A, and SBI can be small and stable, whereas LLM reasoning dominates latency and variability; one measured end-to-end operation took 12.81 s on average, while the underlying SBI query was on the order of 0.49 ms (Garigipati et al., 4 May 2026). This mismatch sharpens the safety question: if action interfaces are fast and precise but reasoning remains slow or hallucination-prone, then validation, rollback, policy checking, and selective autonomy become architectural necessities rather than optional safeguards. Slice orchestration proposals respond by combining MCP with policy and safety agents and reasoning-model governance, but they also leave open the formal verification of LLM-generated manifests, tool calls, and economic decisions (Bandara et al., 27 Jan 2026).

The research agenda is not only technical but organizational. A maturity-model study of agentic AI in 6G software businesses identified 29 motivators and 27 demotivators grouped into five themes each. Motivators included scalable autonomy, cost efficiency, adaptive intelligence, alignment with 6G architecture, and innovation and differentiation; demotivators included technical immaturity, trust and accountability gaps, integration complexity, organizational readiness issues, and cost and performance overheads. That work frames the move to agent-first systems as an organizational and technical transformation grounded in Data, Business Logic, and Presentation layers rather than as a purely algorithmic upgrade (Zohaib et al., 5 Aug 2025).

Taken together, these strands indicate that agentic AI-native 6G is best understood as an architectural direction rather than a settled technology stack. It combines intent-based networking, semantic-native communication, reasoning-native coordination, tool-grounded control, and continual adaptation, but it still faces open problems in interoperability, real-time inference, safety verification, ontology maintenance, large structured action spaces, and the economics of deployment. A plausible implication is that the first durable implementations will be hybrid: deterministic 3GPP infrastructure below, bounded and policy-governed agents above, and simulation or digital-twin validation in the middle.