Industrial Internet of Agents: Conceptual Architecture, Applications, and Challenges

1. Introduction

The Industrial Internet of Things (IIoT) has emerged over the past decade as a core infrastructure for industrial digitalization by enabling large-scale sensing, device interconnection, and data-driven optimization [1,2]. However, as industrial systems become more heterogeneous and interconnected, conventional IIoT remains limited in semantic data utilization, cross-system coordination, and proactive response to unexpected events [3,4,5]. Concurrently, Foundation Models (FMs), especially large language model (LLM)-based agents following paradigms such as ReAct [6], have improved natural-language understanding, tool use, memory, and multi-step reasoning. These capabilities make it possible to move part of industrial automation from fixed rule execution toward intent-driven task planning.

Existing intelligent industrial architectures can be grouped into three lines. Cloud-centric AI analytics support global optimization but suffer from latency, centralization, and limited device-level intelligence [7,8,9]. Edge AI and federated learning improve latency and privacy but provide limited inter-node semantic coordination [10,11,12,13]. IIoT architectures enhanced with LLM agents can handle natural-language intent and support multi-agent negotiation, but often rely on predefined roles or static topologies [14,15,16]. Meanwhile, general Internet-of-Agents studies address scalable collaboration beyond industrial settings [17,18]. Existing surveys mainly cover general agent capabilities or single dimensions of industrial intelligence, while few jointly examine networked agent collaboration and heterogeneous deployment requirements. Although recent work has begun to use the notion of the Industrial Internet of Agents for communication-control reconfiguration [19], a conceptual architecture that connects FM-driven agents with industrial communication, OT interoperability, tiered latency, and verifiable execution remains under-explored.

To help address this gap, this paper conceptualizes the Industrial Internet of Agents (IIoA) and presents a conceptual architecture for analyzing its industrial requirements. As shown in Figure 1, IIoA is derived from the Internet of Agents (IoA) framework of Wang et al. [20], which provides the general scaffold for sensing, communication, decision-making, and execution, but re-specifies it for industrial deployment. Compared with general IoA, IIoA foregrounds operational technology (OT) interoperability, real-time latency, physical safety, execution verification, traceability, human oversight, and cross-enterprise governance, organizing them into constraint-oriented extensions across sensing, communication, decision-making, and execution. The architecture also highlights two deployment tensions: probabilistic FM behavior versus deterministic industrial requirements, and networked agent collaboration versus resource constraints in edge and end environments.

To keep the discussion grounded in current research, this paper synthesizes representative studies connecting AI-agent research with industrial systems. The discussion emphasizes 2023–2026 publications, while retaining earlier works that provide foundational context for IIoT, MAS, communication, and industrial safety. These studies are used as reference points for the conceptual synthesis and for positioning IIoA within existing industrial-intelligence research.

The main contributions are summarized as follows.

1.

This paper defines the Industrial Internet of Agents (IIoA) as a networked industrial-agent concept characterized by FM-driven node cognition, networked collaboration, and end-to-end industrial constraints.

2.

This paper presents an IoA-based conceptual architecture for IIoA, covering hierarchical foundation models and industrial extensions across sensing, communication, decision-making, and execution, and clarifies how networked collaboration is conditioned by OT interoperability, latency, safety, and trust.

3.

Representative work is synthesized across Shopfloor Operations, Autonomous Plant Operations and Maintenance (O&M), Supply Chain Collaboration, and Digital Twin Systems to identify architectural coverage, networking requirements, and engineering bottlenecks.

The remainder of this paper is organized as follows. Section 2 defines IIoA and presents the architectural design. Section 3 discusses recent advances in the context of representative industrial scenarios. Section 4 analyzes the open challenges facing IIoA deployment and identifies directions for future research. Section 5 concludes the paper.

2. Conceptual Architecture

2.1. Concept of IIoA

IIoA extends the lineage of industrial MAS that predates LLMs. Early industrial agents relied primarily on hard-coded rules, statistical algorithms, or task-specific machine learning models to execute predefined tasks. These systems achieved reliable performance in constrained settings, with limited generalization and contextual understanding. Table 1 clarifies the position of IIoA relative to adjacent concepts in industrial intelligence. Accordingly, the Industrial Internet of Agents (IIoA) is defined as a distributed industrial intelligence system in which heterogeneous AI agent nodes are interconnected via network protocols. Given natural-language goals or structured task requests, IIoA supports autonomous sensing, collaborative decision-making, and verifiable execution across nodes. The framework complements IIoT, CPS/CPPS, and industrial MAS through an agent-oriented architectural lens. This lens brings FM-enabled cognition, semantic coordination, and industrial execution constraints into one analytical frame.

Specifically, IIoA has three key characteristics:

1.

FM-driven node cognition. Each agent node may be enabled by or connected to foundation models, supporting independent parsing, planning, and agent-level execution for natural-language goals or structured task requests.

2.

Networked collaboration. Isolated agents cannot support complex industrial workflows. IIoA therefore enables heterogeneous agents to collaborate through dynamic service discovery, cross-node semantic communication, and state-aware multi-turn interaction. Such semantic communication helps reduce integration complexity and supports tighter system integration.

3.

End-to-end industrial constraints. Sensing, communication, decision-making, and execution must all satisfy real-time, physical feasibility, safety, and interoperability constraints, which define the operational and compliance boundaries of both semantic cognition and networked collaboration.

Building on the definitions above, this paper adopts the IoA framework of Wang et al. [20] as its architectural starting point, while re-specifying its sensing, communication, decision-making, and execution modules around industrial constraints. As shown in Figure 1, the following sections identify the industrial limitations of each IoA module and derive the corresponding IIoA extensions, which serve as the basis for the application analysis. The analysis compares the network mechanism, motivating industrial constraint, and available evidence for each extension. It also follows dependencies across modules, since sensing errors, communication delays, planning uncertainty, and execution feedback can propagate through the industrial loop.

2.2. Foundation Models

In the IoA framework, foundation models (FMs) provide the cognitive backbone for sensing, communication, decision-making, and execution. Their deployment model remains open under industrial constraints. For IIoA, model placement directly affects latency, reliability, resource efficiency, and safety, so FM configuration belongs within the system architecture. Throughout this paper, FM or FMs refers collectively to LLMs, VLMs, VLAs, and related model families, while retaining original terminology for specific works.

Industrial missions impose multi-level latency and reasoning requirements that are less prominent in general-purpose deployment. Global planning may tolerate second-level responses, inter-node coordination often requires sub-100 ms interaction, and end-side perception or control may require millisecond-level reaction. A single FM scale cannot satisfy all tasks. Large models support complex reasoning with higher inference latency, while lightweight models respond faster with narrower planning capacity [21,22]. A second constraint is domain adaptation. General-purpose pretraining provides limited coverage of equipment terminology, process specifications, safety constraints, and industrial operating knowledge, so direct deployment may weaken industrial reasoning performance [9]. Table 2 summarizes FM deployment across the three tiers.

To address the heterogeneity of cognitive load, IIoA organizes FMs in a cloud-edge-end tiered architecture. Latency budgets and cognitive complexity guide the partition, with parameter count serving as a secondary consideration. Cloud-tier models handle global intent decomposition and complex cross-node reasoning; edge-tier models support local anomaly diagnosis, task reallocation, and coordination; end-tier models focus on rapid perception and device-level response. Techniques such as quantization, pruning, acceleration, and cloud-edge collaborative inference further support latency-aware deployment [21,22].

To address domain knowledge adaptation, IIoA introduces knowledge injection at each tier through weight post-training, retrieval augmentation, fine-tuning, and cross-tier knowledge distillation. Industrial Foundation Model (IFM) illustrates domain post-training for industrial knowledge injection [23], while cloud-side RAG and edge-side fine-tuning provide lighter mechanisms for adapting general FMs to industrial tasks [9,28]. These mechanisms help bridge gaps between general-purpose FMs and industrial terminology, process rules, and safety logic. Model updates in industrial environments also differ from internet-facing deployment: version transitions should pass offline verification and safety certification before production use.

2.3. Sensing

IoA partitions sensing into modality sensing and functional sensing. These categories cover observable input channels and self-monitoring of reasoning budget, knowledge boundaries, and tool availability. Industrial deployment adds two constraints to this design foundation. Industrial field signals use specialized formats and streaming protocols with engineering-specific semantics. IoA also treats nodes largely in isolation, leaving factory topology outside the sensing context needed for cross-node coordination. Table 3 summarizes the corresponding extensions.

Industrial Signal Awareness. The IoA sensing pipeline targets general modalities such as text, GUI, speech, and video. Industrial perception also depends on vibration, acoustic, thermal, electrical, and other physical signals with distinct spectral structures, sampling densities, and unit semantics. Direct tokenization of numerical streams can weaken temporal continuity and physical meaning. General FMs also show limited alignment with engineering time-series, anomaly patterns, and operating-condition associations [30]. An industrial sensing layer must convert physical streams from the OT Protocol Bridge into representations that preserve temporal and engineering semantics. It must also fuse cross-modal evidence for device-state understanding. Existing studies span successive levels of specialization. Time-LLM explores general time-series reprogramming [29], FD-LLM aligns engineering signals with LLM embedding spaces [30], and QwIndustry combines industrial vibration, acoustic, and visual inputs [31]. The literature therefore progresses from generic sequence representation to domain alignment and multimodal state interpretation. Validation remains concentrated on individual tasks, modalities, and operating regimes.

Physical Context Awareness. IoA sensing is primarily node-centric and provides limited support for factory topology. In IIoA, agents correspond to physical units whose decisions depend on production-line position, process dependencies, and neighboring states. Disturbances can propagate along these dependencies, so topology-blind sensing may classify system-level disruptions as local events [32]. IIoA can represent equipment locations, line structures, and inter-node dependencies as agent-queryable knowledge graphs. Digital twins or world models can update this context when nodes move, fail, or lose sensor coverage. Topology-aware scheduling and digital-twin studies indicate that physical context can improve cross-node coordination [33,34]. Qin and Lu also report sensitivity to changes in machine and job settings [33]. This result places topology maintenance inside the runtime sensing problem. The topology layer must remain synchronized with changing shopfloor configurations and available for agent queries.

2.4. Communication

The IIoA communication module builds on the IoA three-layer semantic stack [20]. The environment layer covers deployment and observation-action interfaces. The interaction layer covers naming, routing, session management, and topology. The semantic layer covers message representation and long-term memory.

Industrial communication adds three engineering constraints to this stack. Proprietary fieldbus protocols separate OT networks from semantic agent networks. Control, coordination, and planning tasks also impose latency requirements that differ by orders of magnitude. Cross-organizational collaboration adds data-sovereignty, identity, and cybersecurity requirements. Table 4 summarizes the resulting Industrial Communication Extensions.

Industrial agent networking also depends on capability discovery. Agents must advertise machine-interpretable capabilities and update them as device state, availability, and permissions change. A two-stage IoA framework combines autonomous capability announcement with task-driven discovery and composition [18]. Its simulation results provide a transferable mechanism for industrial service discovery. Industrial use additionally requires capability records that remain synchronized with equipment state, physical topology, and authorization policies.

OT Protocol Bridge. IoA environments are mainly web, desktop, or simulated systems. Industrial devices communicate through Modbus TCP, OPC UA, and related protocols with distinct frame formats, sampling periods, addressing rules, and safety constraints [35]. IIoA therefore calls for an OT Protocol Bridge that translates fieldbus data into agent-consumable states and converts agent intents into standardized command-write interfaces. Gateway experiments demonstrate millisecond-scale acquisition cycles for ModbusE-to-OPC UA communication [35]. LLM-assisted AAS generation reports effective generation rates of 62–79% under tested configurations [36]. Text-to-OPC UA modeling and LLM-agent control provide additional evidence for exposing device semantics to agent workflows [37,38]. These results establish protocol exposure and model generation as feasible building blocks. End-to-end semantic consistency still requires verification from field signals through agent reasoning to executable commands.

Tiered Latency Design. Industrial tasks occupy several latency classes. Local control requires millisecond responses, edge coordination operates at hundreds of milliseconds, and global planning may take seconds [25]. A uniform message class would expose time-sensitive synchronization to congestion from semantic planning traffic. IIoA therefore stratifies communication by latency budget and task criticality. Control, coordination, and planning traffic use separate service classes while retaining semantic exchange across tiers. CE-CoLLM shows that contextual-data transfer can dominate cloud-edge inference latency [39]. Dual-loop architectures also provide a basis for hierarchical edge-terminal coordination [40]. Evidence is stronger for inference placement than for deterministic multi-agent messaging under OT timing. Industrial deployment must jointly preserve semantic flexibility, bounded latency, reliability, and schedulability.

Security & Trust Design. Cross-organizational collaboration must preserve data sovereignty, verify agent identity, and satisfy cybersecurity requirements. DID-based verification addresses identity across trust domains [41]. Federated and eavesdropping-aware learning address protected model and data exchange [42,43]. IEC 62443 supplies zone-and-conduit governance and lifecycle controls for industrial systems [44]. These studies cover complementary security layers. Integrated policies that connect identity, intent, data access, and executable authority remain scarce. IIoA therefore extends zone protection with controls over what agents may know, infer, request, and execute.

2.5. Decision-Making

The IoA decision-making module comprises task-level planning, uncertainty awareness, and collaborative reasoning. Industrial deployment adds step-wise feasibility checking, task routing across heterogeneous FM tiers, and structured human intervention. Feasibility checks must cover device capabilities, process precedence, and safety exclusion zones during plan generation. Task routing must account for model capacity, latency, and physical criticality. Human engineers also need defined escalation and override interfaces for uncertain or high-risk decisions. Table 5 summarizes these extensions.

Feasibility-Coupled Planning. Industrial plans must satisfy device limits, process precedence, and safety exclusion zones. Errors can propagate along task dependencies when verification occurs after complete plan generation. IIoA therefore embeds constraint solvers and formal checkers into the sense-think-act loop for progressive verification and correction. SMT-based planning and VeriPlan demonstrate solver-backed generation and rule refinement [45,47]. ConstraintLLM evaluates formalization on the 140-task IndusCP benchmark [46]. The evidence supports toolized feasibility reasoning at the planning stage. Plant integration must still reconcile rigorous verification with open-ended task generation and changing operating states.

Tiered Inference Allocation. IIoA’s tiered deployment creates capability asymmetry across cloud, edge, and end agents. Large models support complex reasoning with higher latency, while lightweight agents respond faster with narrower planning capacity. Bidirectional allocation lets upper-tier agents decompose intent and dispatch subtasks. Lower-tier agents return execution states, same-tier agents exchange local evidence, and low-confidence tasks escalate upward. HuggingGPT, HiPlan, and DynTaskMAS provide mechanisms for model routing, hierarchical planning, and dynamic task-graph execution [48,49,50]. Most evidence comes from general-purpose agent settings. Industrial routing must combine model capacity with latency, dependency structure, confidence, and physical task criticality.

Human Oversight Interface. Operations governed by safety or compliance rules require defined conditions for human intervention. IIoA links intervention to confidence and risk through confirmation, guided correction, and override. KnowNo calibrates help requests to uncertainty [51]. General human-in-the-loop studies support guided correction [52]. VeriPlan adds user refinement of verification rules [47]. AgentSpec demonstrates rule-based termination of unsafe actions across several agent domains [53]. These mechanisms remain largely outside industrial deployment. An auditable oversight interface must record the trigger, human response, affected plan, and resumed execution state.

Across the decision literature, formal planning and rule enforcement have the clearest evaluation protocols. Evidence for tiered industrial coordination and risk-triggered human escalation is less mature. A complete decision layer must connect solver outcomes, routing decisions, confidence estimates, and human authority within one traceable workflow.

2.6. Execution

Industrial execution adds three requirements to IoA’s software-oriented design. Device-specific commands need syntax and safety certification before dispatch. Physical failures need staged responses that separate immediate safe-state control, local recovery, and decision-layer escalation. Compliance auditing also requires structured traces that connect reasoning, commands, outcomes, and responsible agents. Table 6 presents the corresponding IIoA extensions.

Semantic Execution Gateway. Industrial execution translates semantic intent into PLC, CNC, AGV, and robotic commands that directly affect physical assets. The execution interface must therefore include syntax compliance and safety verification. A Semantic Execution Gateway converts validated plans into device-specific commands and certifies them before dispatch. PLC workflows may combine IEC 61131-3 Structured Text generation, compilation, formal verification, and iterative repair. CNC and AGV systems require equivalent validation for G-code or motion commands. LLM4PLC demonstrates iterative verification on a manufacturing testbed [54]. Agents4PLC and MPC-Coder extend this direction through closed-loop generation and repair [55,56]. MPC-Coder reports 100% syntactic correctness and 78% functional consistency in its evaluation [56]. These results support a certification boundary between semantic plans and physical control.

Execution Fault Recovery. Physical-device failures can be irreversible and may require responses within millisecond control cycles. IIoA therefore separates immediate safety response from agent-level recovery. Hardware interlocks or safety PLCs first bring the system to a safe state independently of FM inference. Local agents then generate recovery sub-plans and escalate cases that exceed local capability. Recent work combines behavior-tree recovery, ontology-triggered replanning, and constraint monitoring for embodied or industrial agents [57,58,59]. Response time and physical risk determine the recovery stage and escalation path.

Decision Traceability. Industrial deployment requires compliance auditing, incident reconstruction, and accountability for high-risk actions. IIoA binds each executed action to records covering agent identity, reasoning summary, command, timestamp, and outcome. Provenance graphs and runtime rule enforcement support auditability [53,60]. Failure-localization records can support post-incident root-cause reconstruction [61]. Traceability therefore turns execution records into evidence for safety governance and cross-agent accountability. Distributed records still need causal links when several agents contribute to one physical outcome.

Together, the four modules form a coupled industrial loop. The OT Protocol Bridge exposes field states, and sensing converts them into signal and topology context. Communication discovers suitable agents and carries this context across latency tiers. Decision-making allocates reasoning, verifies feasibility, and invokes human authority when required. The execution layer certifies commands, manages recovery, and returns provenance-linked outcomes to subsequent sensing and planning cycles.

This coupling creates recurring design tensions. Semantic exchange supports flexible collaboration, while control traffic requires bounded timing. Cloud reasoning expands global context, while edge and end autonomy protect response time and availability. Adaptive planning also depends on certification before physical execution. Current evidence is strongest for individual building blocks such as protocol gateways, constrained planning, and PLC verification. Dynamic discovery, cross-tier coordination, and full-loop traceability still rely largely on transferable mechanisms or limited prototypes. The application scenarios below show how far these components have progressed toward integrated deployment.

3. Applications

Section 3 examines how the architecture appears in representative industrial deployment contexts. As shown in Figure 2, Shopfloor Operations covers device-level collaboration and execution. Autonomous Plant O&M covers factory networks, computing resources, and equipment management. Supply Chain Collaboration extends coordination across organizations. Digital Twin (DT) Systems also provide shared infrastructure for simulation, safety checking, and pre-execution validation. Across these scenarios, the analysis compares coordination scope, networking role, physical coupling, and validation setting. Representative works are summarized in Table 7.

3.1. Shopfloor Operations

Shopfloor control has traditionally relied on Manufacturing Execution Systems (MES) for order decomposition and centralized dispatching. This architecture is effective in stable, high-volume production, but less adaptable in high-mix, low-volume settings, where disruptions, combinatorial scheduling constraints, and device-level variability require faster local coordination.

Under IIoA, shopfloor entities such as CNC machines, robotic arms, and AGVs can be represented as agent nodes. These nodes parse production intent, announce capabilities, negotiate task allocation, and form temporary collaborative subnetworks according to real-time device states. Existing work can be organized into two lines.

LLM-Based Scheduling. This line focuses on the decision path from production intent to task assignment. MASC, RAG-based scheduling, and evolutionary dispatch-rule generation suggest that LLMs can support dynamic rescheduling, constraint-aware assignment, and localized adaptation under changing shopfloor states [16,24,62]. These studies make scheduling more responsive to real-time context and operational knowledge, although the transition from generated schedules to verified device-level execution remains weakly connected.

Decentralized Agent Collaboration. This line is closer to IIoA’s networked coordination logic. Resource agents negotiate task allocation through bidding, natural language, or G-code-mediated interaction, reducing dependence on centralized schedulers and allowing more flexible shopfloor reconfiguration [14,15]. Compared with scheduling-centered studies, these works shift attention toward distributed coordination, yet protocol-based reassignment under node joining, departure, and failure remains insufficiently tested in real-scale shopfloor environments.

Shopfloor evidence is strongest for schedule generation and negotiation under modeled constraints. Decentralized studies add capability exchange and local task allocation, although their evaluations cover limited node dynamics and physical scale. Verified execution remains the least developed stage. An end-to-end IIoA workflow must connect scheduling and negotiation with certified PLC, CNC, robot, and AGV commands. It must also preserve motion constraints, hard real-time loops, and fault recovery across devices.

3.2. Autonomous Plant O&M

Factory infrastructure management covers industrial networks, edge computing resources, and equipment maintenance. Expert operators have traditionally managed these systems through fixed rules. Growing edge-node counts and high-frequency alarms make manual response and local rule optimization increasingly difficult. IIoA coordinates infrastructure, computing, and equipment agents around operational intent, with engineers retaining supervisory authority. Existing work can be organized along three dimensions.

Network Management. Research on network management explores how LLM agents support optimization, diagnosis, and workflow execution. NetLLM covers multiple network tasks, while NetAssistant exposes diagnostic workflows through natural-language interaction [63,64]. Confucius translates operational intent into verifiable action sequences through specialized agents and DAG workflows [65]. It has operated in Meta’s production network for two years and supports more than 60 applications. This is strong production evidence for data-center network management. Its transfer to industrial OT remains untested under fieldbus interoperability, hard real-time constraints, safety procedures, and plant-level compliance.

Edge Computing Resource Management. Edge resource management focuses on dynamic allocation of computing, communication, and deployment resources under changing production loads. Existing studies cast network slicing, task offloading, and microservice deployment as agent-mediated decision problems, using LLMs or specialized agents to parse intent, decompose constraints, and coordinate resource decisions [66,67,68]. The main deployment issue is whether such negotiation can remain stable and timely when edge nodes frequently join, leave, or shift workloads under real-time industrial constraints.

Equipment Maintenance. Equipment maintenance applies LLM agents to prognostics and health management (PHM), fault diagnosis, and maintenance recommendation. PHM-agent and fault-diagnosis studies provide examples of agents translating operational intent into maintenance workflows and interpreting raw sensor data with natural-language explanations [69,70]. PHM Copilot further grounds maintenance support in standard protocols and copilot-style interaction [71]. The emphasis remains largely device- or task-specific. Collaborative maintenance would let a device agent coordinate detected degradation with scheduling, spare-parts, digital-twin, and execution agents. This coordination could form an autonomous response loop.

Autonomous Plant O&M has uneven evidence across its three functions. Network management includes production-scale data-center deployments, while resource orchestration is evaluated mainly through modeled communication and computing environments. Maintenance studies often use historical data or copilot workflows with human experts. A plant-level IIoA must integrate these functions under OT protocols, real-time constraints, safety procedures, and maintenance authority. Such integration would connect diagnosis, resource adjustment, maintenance planning, and safe execution.

3.3. Supply Chain Collaboration

Traditional supply chain management relies on centralized enterprise resource planning (ERP) and supply chain management (SCM) systems with fixed interfaces. This limits visibility beyond tier-one suppliers, amplifies the bullwhip effect through local optimization, and keeps disruption response largely reactive.

From an IIoA perspective, supply-chain participants can be represented as agent nodes. Semantic exchange can operate above electronic data interchange (EDI) and predefined application programming interfaces (APIs). Cross-enterprise collaboration also faces conflicts of interest, data-sovereignty constraints, and trust barriers. Current studies therefore concentrate on intra-enterprise optimization, public-information sensing, and simplified negotiation settings. Existing work can be organized along three dimensions.

Supply Chain Planning. Planning studies use agents to support demand forecasting, inventory allocation, and replenishment negotiation. LLM-agent negotiation, InvAgent, and SCPA show that agents can improve consensus formation, node-level replenishment decisions, and intent-driven planning across large stock keeping unit (SKU) spaces [72,73,74]. This line moves supply chain planning from fixed optimization routines toward more adaptive semantic coordination, but privacy-preserving cross-enterprise negotiation is still only lightly explored.

Production Logistics Coordination. Logistics coordination focuses on real-time scheduling and response from raw material intake to finished goods delivery. ChatSync links IIoT data, expert knowledge, and spatio-temporal graphs of resources to support production coordination. In a manufacturing case involving air conditioners, it achieved a 91.2% delivery punctuality rate [75]. A related line uses agentic knowledge synthesis to convert unstructured work orders and operational records into structured knowledge bases for retrieval and decision support [76]. Together, these studies suggest that logistics agents depend on both planning algorithms and continuously maintained operational knowledge.

Risk Sensing and Response. Risk-oriented work addresses the limited visibility of multi-tier supplier networks. Supplier knowledge graphs and GenAI-based relationship prediction infer hidden links from public unstructured sources [77,78]. Multi-agent disruption monitoring further maps external signals to cascading risk propagation [79]. The latter framework is evaluated on 30 synthesized disruption scenarios. Current evidence therefore supports external risk sensing and network reconstruction. Reciprocal collaboration among firms remains largely unevaluated.

Supply chain studies extend IIoA beyond the factory through planning, logistics knowledge, and risk sensing. Their evidence is strongest for simulated inventory decisions, single-enterprise logistics, and public-data relationship inference. Cross-enterprise deployment requires agents to negotiate and coordinate while protecting sensitive operational data. Identity, permissions, incentives, and enforceable commitments therefore become part of the networking problem.

3.4. Digital Twin Systems

Digital twins are already used for monitoring, simulation, and predictive maintenance. Their deployment still depends heavily on expert-led model construction, heterogeneous protocol integration, and manual interpretation. Under IIoA, digital twins support agent-assisted modeling, calibration, and querying. They also provide virtual infrastructure for simulation, safety checking, and pre-execution validation. This section discusses agent-driven twin construction and twin-supported agent reasoning.

Agent-Driven Twin Modeling. This direction uses agents to reduce manual effort in digital twin construction and operation. LLM multi-agent calibration shows how agents can explore parameter spaces and inject model prior knowledge into DT parameterization [80]. More broadly, agent-DT integration studies point to a shift toward agent-assisted twin workflows [81]. CNC-oriented RAG querying and MTConnect-based legacy-device integration further show how agents can provide semantic access to machine states and specialized supervision over heterogeneous equipment [7,82]. The main shift is from manually configured twin models to agent-assisted twin services, though current systems are still mostly validated on specific machines, protocols, or modeling tasks.

DT-Driven Agent Reinforcement. Digital twins provide resettable environments for agent training and strategy evaluation, reducing the cost and risk of production exploration. Collaborative DT training, inverse-reinforcement-learning (RL) trajectory optimization, and dual-twin transfer learning provide mechanisms for reinforcement and generalization [83,84,85]. Several studies target conventional learning agents. Their results still inform policy learning, failure memory, and sim-to-real adaptation within IIoA.

DT-Driven Reasoning Simulation. DTs can serve as pre-execution sandboxes where agents test strategies before physical dispatch. Process-plant and reactor studies combine twin knowledge bases, simulation tools, OPC UA invocation, and formal constraints [86,87]. These components support an agent-generate, twin-verify, physically-execute loop. Current demonstrations focus on individual systems or bounded agent workflows. Shared DT infrastructure must also maintain state consistency, data freshness, model validity, and access control across multiple agents.

Digital-twin evidence spans laboratory machine integration, simulation-based learning, and constrained control evaluation. Gautam et al. connect legacy and smart machines through MTConnect and specialized agents [7]. The system demonstrates physical data access and bounded diagnostic functions, while broader machine coverage and composite multi-agent queries remain future work. DTs therefore provide the clearest bridge between virtual reasoning and physical context. Their value as shared IIoA infrastructure depends on synchronized models and validated sim-to-real transfer.

Across the four scenarios, application evidence forms an uneven maturity profile. Shopfloor and DT studies provide the strongest physical or testbed coupling, but usually within bounded tasks and small system scopes. Plant O&M includes production-scale agent deployment in data-center networks, with industrial OT transfer still open. Supply-chain work relies more heavily on simulation, synthesized disruptions, and public information. The literature concentrates on signal awareness, physical context, tiered inference, and feasibility planning. Security and trust, human oversight, execution fault recovery, and decision traceability receive much less application-level validation. This imbalance defines the transition from task-specific industrial agents to an integrated IIoA system.

4. Challenges and Outlook

The preceding sections introduced IIoA’s conceptual architecture and related research progress. Large-scale industrial deployment, however, faces unresolved tensions that span multiple modules. Table 8 summarizes seven challenge dimensions: execution reliability, inference latency, OT protocol integration, collaboration scalability, physical foundation, continuous adaptability, and evaluation.

Across these challenges, networking is the mechanism that couples agent cognition to industrial execution. Semantic messages must coexist with deterministic control traffic, while agent discovery, routing, and coordination must remain stable as edge nodes join, leave, or fail. This creates a need for cross-layer evaluation that links communication delay and jitter to reasoning quality, task completion, and physical safety. Future benchmarks should report task success together with protocol-translation fidelity, communication overhead, tail latency, coordination convergence, churn resilience, and recovery time. Reproducible OPC UA and Modbus testbeds, combined with digital-twin sandboxes, can provide controlled settings for tracing failures across the sensing–communication–decision–execution pipeline. Security and trust also need evaluation at the interaction level. Identity and zone protection remain incomplete when authenticated agents can request unauthorized information or issue semantically unsafe commands. Communication, inference placement, trust governance, and execution verification therefore form a coupled set of design variables.

Progress on these coupled design variables requires early coordination among AI researchers, industrial-control experts, and standards bodies. The coordination agenda includes agent identity, OT protocol semantics, execution traceability, formal safety constraints, and industrial semantic alignment. Early agreement can reduce fragmentation as technical approaches mature.

5. Conclusions

This paper introduces the Industrial Internet of Agents (IIoA) as a concept for networked industrial agents and presents a conceptual architecture for analyzing industrial requirements. The synthesis places agent cognition, network coordination, OT interoperability, and verifiable execution within a jointly designed industrial stack. Progress remains uneven across shopfloor operations, plant O&M, supply-chain collaboration, and digital twin systems. The gap becomes most visible when capabilities from isolated or controlled settings move into heterogeneous industrial networks. Grounded in representative literature, the paper contributes an analytical framework and a cross-scenario synthesis of current evidence. Empirical evaluation remains a priority for subsequent work. End-to-end prototypes, shared testbeds, and longitudinal operational evidence should examine protocol-translation fidelity, latency, coordination scalability, trust governance, execution reliability, and recovery from node or communication failures.