Agentification of Edge Intelligence

• Agentification of edge intelligence is defined as transforming passive edge nodes into autonomous cognitive agents enabling multi-task autonomy and distributed decision-making.
• It leverages techniques like adaptive Chain-of-Thought, sparse expert activation, and world-model planning to optimize latency, energy, and resource usage.
• Studies show that agentic edge systems improve real-time operations in applications such as vehicular networks, IoT, and industrial automation with measurable performance gains.

Agentification of edge intelligence denotes the transformation of passive edge nodes into autonomous cognitive agents that execute continuous perception–reasoning–action loops, collaborate with peer agents, and operate under stringent latency, energy, memory, privacy, and connectivity constraints. In the recent literature, closely related formulations include Edge General Intelligence (EGI), where distributed agents perceive, reason, and act autonomously across dynamic environments, and Mobile Edge General Intelligence (MEGI), which emphasizes real-time, privacy-preserving reasoning at the wireless network edge (Zhang et al., 26 Aug 2025, Luo et al., 27 Sep 2025). Across these formulations, the unifying theme is that edge intelligence ceases to be a static inference service and becomes a distributed agentic system with explicit reasoning, planning, adaptation, and coordination.

1. Conceptual foundations and formalizations

Agentification is defined in several mutually compatible ways. In EGI, it is the process of transforming edge-level sensing–compute–actuation units into autonomous, cognitive agents capable of perception, reasoning, and action, typically by equipping devices with an LLM core, perception modules, reasoning modules, and action interfaces (Liu et al., 27 Aug 2025). In another formalization, each edge agent is modeled as an MDP

$A = (S, A, P, R, \pi),$

where the state space captures multimodal observations, the action space includes decisions such as offloading, beamforming, or waypoint control, and the policy is learned or refined at the edge (Zhang et al., 26 Aug 2025). These formulations are complemented by a functional decomposition in which an agent is written as

$$A = \bigl(\mathcal{P},\;\mathcal{L},\;\mathcal{R},\;\mathcal{C}\bigr),$$

with perception, LLM-based cognition, reasoning/planning, and control interfaces as distinct components (Liu et al., 27 Aug 2025).

A central distinction in the surveys is between traditional edge intelligence and agentic edge intelligence. Traditional systems are described as single-task, inference-driven, and largely static; agentified systems instead feature multi-task autonomy, online policy updates via RL or continual learning, explicit planning or CoT-style reasoning, decentralized multi-agent coordination, richer memory, and joint energy–latency–QoE optimization (Zhang et al., 26 Aug 2025). This distinction is not merely terminological. It redefines the edge node from an endpoint that executes a precompiled model into an adaptive decision-maker embedded in a broader MAS.

A more proactive formalization appears in the world-model literature, which treats the edge agent’s environment as an underlying POMDP and learns latent-state encoders, dynamics priors, observation decoders, and reward models. Planning then optimizes imagined cumulative reward in latent space rather than reacting only to current observations (Zhao et al., 13 Aug 2025). This extends agentification from explicit reasoning over text or tool traces to foresighted internal simulation.

2. Core mechanisms for making edge systems agentic

Recent work identifies several mechanisms through which agentification is realized at the edge. In MEGI, reasoning enhancement begins with methods such as Chain-of-Thought prompting, Supervised Fine-Tuning, and Mixture of Experts, but the key contribution is a joint optimization of reasoning quality and resource efficiency:

$$\min_{\theta,d} \; \mathcal{L}_{reason}(\theta,d)\;+\;\lambda\,\mathcal{C}_{resource}(\theta,d)$$

subject to resource, latency, and accuracy constraints. Here, $\theta$ includes parameters such as expert weights and quantization levels, and $d$ denotes reasoning depth, i.e., the number of CoT steps (Luo et al., 27 Sep 2025).

Within this formulation, adaptive CoT replaces a fixed reasoning budget with a request-dependent depth chosen from task complexity and device capacity. Complex tasks receive deeper reasoning chains, whereas resource-poor devices or lightly loaded nodes use shallower reasoning. The same framework couples adaptive CoT with a distributed MoE architecture in which a BS Control Unit hosts the gating network and core LLM trunk layers, while a pool of edge devices hosts sparsely activated expert subnetworks. For token embedding $x\in\mathbb{R}^d$, expert selection is governed by

$$g_i(x) \;=\; \frac{\exp\bigl(w_i^\top x\bigr)}{\sum_{j=1}^N \exp\bigl(w_j^\top x\bigr)},$$

followed by top-$k$ activation and weighted aggregation $\sum_i g_i(x)\,E_i(x)$ (Luo et al., 27 Sep 2025). In this design, agentification is inseparable from sparse execution and runtime scheduling.

A second line of work emphasizes world models as the “cognitive backbone” of EGI. An edge agent is decomposed into latent representation learning, dynamics modeling, and imagination-based planning modules; planning can be realized through MPC with CEM, MCTS with learned policy/value heads, or actor–critic learning on imagined rollouts (Zhao et al., 13 Aug 2025). This architecture enables anticipatory behavior in vehicular networks, UAV swarms, IoT caching and routing, and NFV orchestration. This suggests that agentification is not limited to LLM-style verbal reasoning; it also includes compact predictive simulators that support multi-step planning under uncertainty.

A third mechanism appears in wireless context engineering, where context itself becomes a controlled resource. The Wireless Context Communication Framework defines a three-stage pipeline—Context Construction, Context Transmission, and Context Inference—and formulates modality selection as a constrained maximization under latency, energy, and memory budgets (Zhao et al., 7 Feb 2026). Rather than increasing model scale, the system agentically selects, filters, structures, and ages context at inference time. This turns “what the model sees” into an active decision variable.

3. Communication, orchestration, and distributed system structure

Because edge agents rarely operate in isolation, agentification also requires communication and orchestration layers. A detailed case study of the Agent2Agent protocol divides communication requirements into system management and information transportation. System management includes agent description and identification, publication and discovery, and orchestration hooks for lifecycle operations and resource allocation. Information transportation includes JSON-RPC 2.0 and modality-agnostic payloads, synchronous request/response, asynchronous SSE and webhook callbacks, and HTTP/HTTPS over TLS as the main wire protocol (Duan et al., 17 Aug 2025). The same study concludes qualitatively that A2A is excellent at heterogeneity, provides a good foundation for scalability and dynamicity, and offers poor support for resource constraints.

An industrial realization of agentification is given by an Industry 5.0 framework in which every logical function is a first-class agent. Its architecture uses an MQTT broker for real-time pub/sub messaging and RESTful HTTP for on-demand calls to external generative AI services. The system defines a Config Loader, CSV Reader, Sensor Streaming, Inference Agent, UI Agent, GenAI Agent, and Designer Agent, each with a sharply delimited I/O contract and one well-defined responsibility (Martinez-Gil et al., 29 Oct 2025). In this formulation, agentification is identical with modular, message-driven decomposition: data ingestion, inference, user interaction, explanation, and deployment are distinct agents rather than stages in a monolith.

Multi-agent coordination is formalized differently in secure multi-LLM EGI. A three-tier heterogeneous deployment places cloud LLMs and policy engines above edge-server LLMs and on-device LLM agents; coordination may follow a pipeline pattern in which a coordinator splits a task into subtasks and routes them securely, or a voting-based consensus pattern in which peer agents broadcast signed and encrypted proposals and decide by majority vote under policy checks (Liu et al., 27 Aug 2025). This design makes inter-agent message-passing a core part of cognition rather than an auxiliary transport function.

EdgeAgentX presents another orchestration pattern, combining FL, MARL, and adversarial defense in military communication networks. Local edge nodes run actor–critic agents with decentralized execution, periodically synchronize through federated aggregation, and receive secure global model updates after anomaly filtering (Ray, 24 May 2025). In this case, agentification extends beyond LLM mediation to distributed policy learning and secure coordination in non-stationary, adversarial environments.

4. Empirical trade-offs: reasoning quality, latency, energy, and failure modes

Empirical studies consistently reject the idea that edge-agent quality is a simple function of parameter count. A domain-conditioned benchmarking study on ITBench defines per-domain accuracy, mixed-domain accuracy, and failure-mode rates for FinOps and SRE tasks, and reports that top FinOps accuracies reach 56–63% whereas top SRE accuracies linger at 6–10% (Wang et al., 11 May 2026). The same study shows a family-agnostic Pareto frontier in the accuracy–latency plane and reports that Qwen Coder 7 B and 32 B both achieve 33.9% overall accuracy, but at $\tau=1.45$ s and $$A = \bigl(\mathcal{P},\;\mathcal{L},\;\mathcal{R},\;\mathcal{C}\bigr),$$0 s, respectively, while Phi4 Mini 3.8 B attains 28.6% at approximately 27 s (Wang et al., 11 May 2026). The practical implication is that model selection at the edge must be domain-conditioned and budget-constrained rather than size-maximizing.

Failure-mode analysis sharpens this point. The benchmarking study distinguishes semantic failures from execution failures such as tool errors and max-steps exhaustion. Qwen is described as semantic-failure-dominated, whereas Phi and Mistral exhibit many execution failures on SRE (Wang et al., 11 May 2026). This matters operationally: when semantic failures dominate, post-verification can be effective; when execution failures dominate, the workflow itself must be simplified or guarded by fallbacks.

MEGI experiments provide a complementary systems view. In a simulated environment with 15 mobile nodes, a BS with 24 GPUs, and devices with 2 TFLOPS each, the proposed MoE + Dynamic CoT configuration achieves 92% latency satisfaction, 95% accuracy satisfaction, and 65 kJ total energy consumption per 1000 requests, compared with 60%, 70%, and 100 kJ for a dense baseline without CoT (Luo et al., 27 Sep 2025). Local validation on Gemma-2B variants reports that Gemma-2B-IT improves accuracy by 15% relative to base, reduces latency by 10%, and that CoT adds transparency but increases latency by 20% (Luo et al., 27 Sep 2025). These results support the narrower claim that sparse expert activation and adaptive reasoning depth can improve the latency–accuracy–energy operating point in resource-constrained MEGI deployments.

Industrial measurements show that agentification can also affect deployment operations, not only inference. In the cream-cheese production case, deployment time shrank by approximately 80% compared to manual containerization and flashing; inference latency averaged under 200 ms on Raspberry Pi 4 and under 150 ms on ESP32 S3; predictive accuracy exceeded 95%; end-to-end system downtime decreased by approximately 65%; and overall energy efficiency improved by approximately 20% (Martinez-Gil et al., 29 Oct 2025). This suggests that modular agent decomposition can compress engineering and maintenance loops as well as runtime loops.

5. Domain-specific manifestations in wireless and edge systems

In 6G Native-AI RAN, agentification is increasingly paired with semantic-native communication. A unified taxonomy classifies work along semantic abstraction level from $$A = \bigl(\mathcal{P},\;\mathcal{L},\;\mathcal{R},\;\mathcal{C}\bigr),$$1 to $$A = \bigl(\mathcal{P},\;\mathcal{L},\;\mathcal{R},\;\mathcal{C}\bigr),$$2, agent autonomy and coordination granularity from $$A = \bigl(\mathcal{P},\;\mathcal{L},\;\mathcal{R},\;\mathcal{C}\bigr),$$3 to $$A = \bigl(\mathcal{P},\;\mathcal{L},\;\mathcal{R},\;\mathcal{C}\bigr),$$4, and RAN control placement from $$A = \bigl(\mathcal{P},\;\mathcal{L},\;\mathcal{R},\;\mathcal{C}\bigr),$$5 to $$A = \bigl(\mathcal{P},\;\mathcal{L},\;\mathcal{R},\;\mathcal{C}\bigr),$$6 (Feng et al., 4 Dec 2025). In this framework, agentification is not only about autonomous policies but about aligning meaning exchange, agent hierarchy, and control-loop placement across PHY/MAC, near-RT RIC, non-RT RIC, and cross-layer orchestration.

Representative use cases make this convergence concrete. For immersive XR, the semantic layer extracts scene semantics into 512-dim tokens, yielding a semantic compression ratio of approximately $$A = \bigl(\mathcal{P},\;\mathcal{L},\;\mathcal{R},\;\mathcal{C}\bigr),$$7 versus raw video, with end-to-end latency around 4 ms and TSR approximately 92% at SNR = 0 dB (Feng et al., 4 Dec 2025). In vehicular V2X, vehicles exchange 256 B intent embeddings of planned trajectories rather than 20 kB raw LiDAR scans, corresponding to $$A = \bigl(\mathcal{P},\;\mathcal{L},\;\mathcal{R},\;\mathcal{C}\bigr),$$8 compression, while TSR gains are reported as +18% at 2 dB and +9% at 5 dB relative to a bit-centric baseline (Feng et al., 4 Dec 2025). In industrial digital twins, hierarchical control across $$A = \bigl(\mathcal{P},\;\mathcal{L},\;\mathcal{R},\;\mathcal{C}\bigr),$$9–$$\min_{\theta,d} \; \mathcal{L}_{reason}(\theta,d)\;+\;\lambda\,\mathcal{C}_{resource}(\theta,d)$$0 achieves 30% latency reduction and 20% energy saving versus semCom-only (Feng et al., 4 Dec 2025).

Other surveys report domain-specific gains from agentic methods in low-altitude economy networks, intent-driven networking, vehicular edge computing, and human-centric service provisioning. LLM-enhanced RL reward shaping in UAV-IoT data collection reduces final energy by 4% for DDPG and 6.4% for TD3, while LLM-TD3 converges 30% faster (Zhang et al., 26 Aug 2025). Agentic Contextual Retrieval raises intent-handling accuracy from 78% to 85% on low-complexity cases and from 65% to 79.8% on high-complexity cases, while reducing communication by 15% and 23.4%, respectively (Zhang et al., 26 Aug 2025). In embodied vehicular MEC, agentic AI improves cumulative return by 61% and converges 45% faster than pure PPO; in human-centric service provisioning, normalized QoE rises from 0.73 to 0.93 (Zhang et al., 26 Aug 2025). These examples indicate that agentification is increasingly treated as a general systems paradigm across wireless control, service composition, and cyber-physical automation.

6. Security, controversies, and open directions

The collaborative structure of agentified edge systems widens the attack surface, particularly in multi-LLM deployments. A zero-trust survey identifies intra-LLM risks such as jailbreaks, prompt injection, memorized data leakage, and toxic outputs, and inter-LLM risks such as chain compromise, insecure communication, unauthorized actuation, Byzantine consensus manipulation, and cross-context data leakage (Liu et al., 27 Aug 2025). Its response is a zero-trust framework built on “never trust, always verify,” least privilege, and continuous monitoring with micro-segmentation. Model-level mechanisms include strong identity and authentication, context-aware access control, and TEE-based integrity verification; system-level mechanisms include proactive maintenance, blockchain-based management, and secure update protocols (Liu et al., 27 Aug 2025).

Communication protocols remain a major bottleneck. The A2A case study argues that centralized registries can become single points of failure, that JSON-RPC over HTTPS can be a bandwidth and latency burden on low-powered devices, and that current agent cards do not yet standardize host-level resource descriptors such as CPU, memory, and battery (Duan et al., 17 Aug 2025). Open issues therefore include resource-aware protocol extensions, decentralized discovery, session handover under mobility, predictive pre-fetching of agent metadata, and virtualized multi-tenant communication fabrics (Duan et al., 17 Aug 2025).

A broader controversy concerns where the executive substrate of an agent should live. A position paper frames the current transition as a “Prefrontal Turn,” arguing that useful agentic intelligence depends more on framework-level executive control than on additional posterior-scale pre-training, and that such control must remain close to local context to preserve cognitive alignment (Tian et al., 18 May 2026). The same paper introduces the “Data-Geography Paradox,” according to which local file hierarchies, telemetry streams, transient OS state, and interaction traces lose fidelity or meaning when prepared for cloud transmission, and the “Interaction-Alignment Loop,” in which implicit user feedback becomes a high-fidelity on-device signal for refinement (Tian et al., 18 May 2026). These claims are programmatic rather than universally settled, but they articulate a widely discussed rationale for edge-native agency.

Open research directions recur across the literature: secure multi-party computation and differential privacy for expert parameters in transit, malicious-node detection, support for heterogeneous accelerators and quantization regimes, decentralized peer-to-peer expert consensus, multimodal reasoning over vision, audio, and sensor streams, safety guarantees for world-model planning, lightweight cryptography for real-time verification, and standardized benchmarks and APIs for cross-platform deployment (Luo et al., 27 Sep 2025, Zhao et al., 13 Aug 2025). Collectively, these directions indicate that agentification of edge intelligence is moving from a model-centric question—how to fit a larger reasoner onto a smaller device—to a systems question concerning distributed cognition, resource-aware orchestration, semantic communication, and verifiable autonomy.