Agentic Vehicles (AgVs): Intelligent Systems

• Agentic Vehicles (AgVs) are intelligent mobile systems that combine autonomous control with goal-directed reasoning and adaptable interactions.
• They use a layered architecture that integrates sensing, cognitive processing, communication, and digital twin components to enhance real-time safety and coordination.
• AgVs are applied across domains—from industrial manufacturing to UAVs—improving efficiency while raising unique security, governance, and operational challenges.

Agentic Vehicles (AgVs) are vehicles that integrate agentic AI to reason, adapt, and interact within complex environments, extending beyond the task automation characteristic of traditional autonomous vehicles. In this usage, agency denotes goal-directed reasoning, contextual adaptation, dialogic interaction, tool use, and lifecycle intelligence, rather than merely the absence of direct human control. The concept is applied to manually driven and autonomous road vehicles, industrial AGVs in factories and warehouses, embodied UAVs, and fleet-level electric mobility operations, making AgVs a cross-domain category of intelligent, mobile systems embedded in larger human-machine ecosystems (Yu, 7 Jul 2025, Eslami et al., 18 Dec 2025).

1. Conceptual foundations

The central conceptual distinction is between autonomy and agency. In the systems-level AgV framework, autonomy refers to operating according to internal rules without external control and to executing preprogrammed tasks independently of external input, whereas agency is the capacity to initiate action based on internal representations of goals, values, or purposes, often in coordination or negotiation with others. On this view, autonomous vehicles can sense, plan, and control without necessarily exhibiting higher-order cognitive functions, while agentic vehicles add goal formation and revision, contextual and ethical reasoning, social interaction, and tool invocation (Yu, 7 Jul 2025).

This distinction is treated as orthogonal to SAE automation. One framework introduces a five-level developmental scale ranging from non-agentic to reflective agentic, while a security-oriented analysis adopts a six-level “agency” capability taxonomy from reactive tools to general agents. The coexistence of these scales indicates that agency is being treated as a separate analytical axis from driving automation, rather than as a simple extension of it (Yu, 7 Jul 2025, Eslami et al., 18 Dec 2025).

A common misconception is that AgVs are merely higher-SAE autonomous vehicles. The literature instead frames them as systems whose behavior is shaped by memory-based personalization, goal interpretation, strategic reasoning, and tool-mediated assistance, with agency appearing in both manually driven and autonomous platforms. This suggests that the defining issue is not whether a human remains in the loop, but whether the vehicle stack can reinterpret goals, coordinate with external actors, and adapt its behavior over longer horizons (Eslami et al., 18 Dec 2025).

2. Architectural organization

A recurring architectural pattern is a layered stack that augments the classical perception-planning-control pipeline with cognitive and communicative mechanisms. One systems-level formulation specifies five interrelated layers: a perception and sensing layer for real-time environmental data acquisition and mapping; a cognitive layer for planning, prediction, and ethical reasoning aligned with dynamic goals and values; an interaction layer for natural language and multimodal exchanges; an execution layer for low-level vehicle control; and a tool interface layer for APIs, infrastructure, and external services (Yu, 7 Jul 2025).

Industrial implementations instantiate this layering in more concrete forms. In a converted industrial vehicle for the primary aluminum industry, the architecture is split between an Autonomous Driving Processing Unit and a PLC-based system supervisor. The software stack is organized into sensing, planning, and control blocks, while manual override is enforced through a finite-state machine with Manual State, Waiting State, and Autonomous State; at ignition the system is always in Manual State, and any manual command activation triggers an immediate reset to Manual State (Iob et al., 2024).

Digital-twin-assisted AGV control in flexible manufacturing expresses a related hierarchy through a physical layer, a first control layer for local maneuvers, a second control layer for global path planning, and a graph-based digital twin with added-value services. The authors explicitly advocate a hybrid digital twin in which some control functions, such as second-layer path planning, can run in the twin while physical safety and some maneuvering functions remain on the vehicle. This division allocates real-time safety-critical behavior to embedded control while moving routing, monitoring, and service logic into a supervisory digital representation (Azangoo et al., 2021).

Cloud-connected AGV architectures sharpen the same split. A generalized cloud architecture distinguishes primary computing needs, such as obstacle avoidance, real-time navigation, and immediate collision avoidance, from secondary computing needs such as daily training of ML models, driving profile storage, and traffic data processing. The reported task finish time of over $300\,\text{ms}$ across evaluated cloud services prevents their use in real-time AGV operation, so the fast, safety-critical agent remains onboard while slower learning and fleet-level analytics are offboard (Damacharla et al., 2020).

3. Shared world models and communication substrates

Many AgV proposals treat shared state as a first-class systems problem. In blockchain-enabled AGVs, each vehicle maintains a local blockchain with blocks of the form

$$H_i=\text{SHA256}(i \Vert H_{i-1} \Vert t_i \Vert D_i),$$

where $D_i$ contains vehicle ID, speed, coordinates, and direction. The design uses a distributed ledger of per-vehicle chains, tied together through a double linked list, and overlays a Named Data Networking access layer so that AGVs request content by name rather than by address. The implementation is a simplified blockchain with no proof-of-work and no explicit consensus algorithm, so it should be understood as a tamper-evident shared memory rather than a full consensus ledger; nevertheless, the study reports higher per-AGV packet transmission and throughput than AODV, with up to $\sim 15\%$ improvement in packet transmission for some AGVs (Liu et al., 2023).

Graph-based digital twins provide a second, more formal shared model. In flexible manufacturing, the factory floor is divided into square-shaped zones $Z_{ab}$, each zone becomes a node, and adjacency is encoded in an adjacency matrix $H_{abcd}$ whose diagonal entries $H_{abab}=W$ model waiting times and whose row and column zeroing removes obstacle zones from the traversable graph. The digital twin is continuously updated as AGVs and infrastructure sensors produce new maps and obstacle detections, allowing path planning and traffic management to operate on a live formal representation of the environment (Azangoo et al., 2021).

Communication quality itself can become part of the shared operational model. In the integrated 5G/industrial digital-twin case study, AGVs periodically exchange positions over 5G using a 5G Network AAS and per-vehicle UE AASs interconnected with the factory model through OPC UA. If communications fail, AGVs stop for safety reasons after $3$ consecutive missed position messages, and resume only after $5$ consecutive successful updates over $5$ seconds; under one reported setting, moving the gNB reduced total plant production loss from $$H_i=\text{SHA256}(i \Vert H_{i-1} \Vert t_i \Vert D_i),$$0 to $$H_i=\text{SHA256}(i \Vert H_{i-1} \Vert t_i \Vert D_i),$$1. This suggests that, in networked AgVs, communication reliability is not a background assumption but an explicit state variable shaping vehicle behavior and plant productivity (Cañete-Martín et al., 16 Jun 2026).

4. Planning, scheduling, and coordination algorithms

The algorithmic landscape of AgVs ranges from decentralized resource protocols to centralized optimization and reinforcement learning. A foundational decentralized protocol models an industrial plant as a graph $$H_i=\text{SHA256}(i \Vert H_{i-1} \Vert t_i \Vert D_i),$$2, defines micro resources as nodes and macro resources as ordered sequences of shared nodes, and equips each AGV with a finite-state machine over REQUEST, WAIT, MOVE, and REPLAN. Coordination occurs through local virtual sign-boards containing fields such as ID, priority, status, current node, next node, and timer; under the stated sampling-time condition, the protocol guarantees mutual exclusive access to contested resources, and a conservative global condition $$H_i=\text{SHA256}(i \Vert H_{i-1} \Vert t_i \Vert D_i),$$3 is given to ensure the absence of deadlocks during system evolutions (Marino et al., 2011).

For fleets executing MAPF plans, dynamic re-ordering can be formulated more explicitly. The Switchable Action Dependency Graph replaces a fixed Action Dependency Graph with a family of switchable inter-agent dependencies, and a comparatively low-dimensional MILP repeatedly selects which ordering to enforce at contested resources. The formulation is recursively feasible and maintains deadlock-free guarantees while dynamically minimizing the cumulative route completion time of all AGVs, thereby turning static MAPF execution into an online coordination problem with formal safety guarantees (Berndt et al., 2023).

Learning-based approaches move coordination into policy learning. In decentralized multi-AGV task allocation, MADDPG-IPF and BiCNet-IPF use an Information Potential Field reward to provide stepwise rewards and implicitly guide AGVs to different material targets under partial observability; compared with baseline methods, the reported gains reach up to $$H_i=\text{SHA256}(i \Vert H_{i-1} \Vert t_i \Vert D_i),$$4 task response improvement and $$H_i=\text{SHA256}(i \Vert H_{i-1} \Vert t_i \Vert D_i),$$5 training iterations reduction. In a separate line of work, Colored-Timed Petri Nets combined with actor-critic model-based reinforcement learning use dynamic action masking and a lookahead strategy for optimal positioning of AGVs, matching traditional methods on smaller instances and outperforming them on larger ones in makespan while achieving a tenfold reduction in computation time (Li et al., 2021, Lassoued et al., 8 Jan 2026).

Scheduling also remains an optimization problem in the classical sense. Hybrid quantum-classical computation for AGV scheduling maps a realistic factory scheduling problem to a railway-like timetabling formulation and reports that a scenario involving up to $$H_i=\text{SHA256}(i \Vert H_{i-1} \Vert t_i \Vert D_i),$$6 AGVs can be efficiently addressed by a hybrid solver in seconds. At the same time, the authors note that the precise contribution of quantum computation remains unclear because the solver is proprietary. This is an example of a broader pattern: AgV coordination is increasingly treated as a heterogeneous optimization stack in which MILP, RL, heuristic scheduling, and hybrid solvers coexist rather than exclude one another (Śmierzchalski et al., 2023).

5. Operational domains and representative systems

Industrial manufacturing remains the most densely developed AgV domain. In flexible manufacturing, AGVs are central mobile resources whose missions can be changed “in less than a second,” and the Aalto Factory of the Future demonstrator uses MiR100 and Seit100 platforms to generate maps, construct graph-based twins, route vehicles with Dijkstra’s algorithm, compute waiting zones and collision areas, and trigger battery charging missions as value-added services (Azangoo et al., 2021).

Warehousing extends the same logic into hybrid labor settings. In a dynamic picker-to-parts warehouse, Neural Approximate Dynamic Programming coordinates a mixed team of human and AGV workers through non-myopic decision making, order batching, and battery management. In the reported baseline with $$H_i=\text{SHA256}(i \Vert H_{i-1} \Vert t_i \Vert D_i),$$7 humans and $$H_i=\text{SHA256}(i \Vert H_{i-1} \Vert t_i \Vert D_i),$$8 AGVs, the learned policy improved filled orders by $$H_i=\text{SHA256}(i \Vert H_{i-1} \Vert t_i \Vert D_i),$$9 relative to Myopic-ILP; in a $D_i$0 humans and $D_i$1 AGVs setting, the gain over a heuristic reached $D_i$2. A different manufacturing study combines a traditional deadlock-safe scheduler with an ANN predictor of future tasks and reports efficiency improvements of up to $D_i$3 to $D_i$4, illustrating a transition from purely reactive dispatch to predictive fleet positioning (Dehghan et al., 2023, Li et al., 2019).

Ports and large outdoor logistics environments emphasize heavy-duty transport and supervisory coordination. A container-terminal Multi-AGV System is proposed around electric AGVs with total motor power between $D_i$5–$D_i$6 and speed of $D_i$7, free ranging navigation using GPS, digital compass and rotary encoder sensors, and a supervisory controller based on a priority based cooperative control algorithm. Here, “agentic” behavior remains strongly hierarchical: vehicles execute local motion and report state, while fleet-level coordination is centralized (Saputra et al., 2021).

The concept also extends beyond ground intralogistics. AirVista-II is presented as an end-to-end agentic system for embodied UAVs, with a Planning Module that identifies temporal scope and modality, and an Execution Module that dispatches image, short-video, or long-video agents for semantic understanding under a zero-shot setting. In electric mobility, an agentic aggregator framework for electric bus fleets couples an optimization-based scheduling model with supervisory agents for disturbance detection, tariff adaptation, and schedule evaluation, while accounting for service reliability, battery state-of-charge, charger availability, electricity prices, route-energy uncertainty, and vehicle-to-grid opportunities (Lin et al., 13 Apr 2025, Manzolli et al., 24 Jun 2026). This suggests that the AgV concept is not tied to a single embodiment but to a recurring organizational pattern: embodied mobility, supervisory reasoning, and explicit coupling to shared digital infrastructure.

6. Safety, security, and governance

AgV research repeatedly separates high-level agency from hard safety enforcement. In the industrial vehicle with manual override capability, the decisive architectural fact is that autonomy is always subordinate to supervisory logic: the PLC-based system supervisor controls mode switching, manual inputs remain armed even in Autonomous State, and any activation of brake pedal, throttle pedal, or steering wheel causes an immediate transition back to Manual State. The reported manual override response time is governed primarily by PLC computation time and communication bus speeds, and remains stable across changes in higher-level autonomy algorithms (Iob et al., 2024).

Communication and computation constraints reinforce the same point. In network-aware AGV control, packet reception ratio is more consequential than pure delay: for $D_i$8, tracking error drops rapidly toward $D_i$9, whereas with $\sim 15\%$0, increasing delay from $\sim 15\%$1 to $\sim 15\%$2 raises the path-tracking error only from $\sim 15\%$3 to about $\sim 15\%$4. Likewise, public cloud services show task finish times of over $\sim 15\%$5, preventing their use in real-time AGV operation. The technical implication is straightforward: real-time safety remains a local systems problem even when planning, learning, and coordination become more agentic (Bragato et al., 8 Sep 2025, Damacharla et al., 2020).

Security analysis further complicates the picture. A role-based AgV architecture with a Personal Agent, a Driving Strategy Agent, and a deterministic Safety Check is proposed precisely because reasoning-driven components create new vulnerabilities such as memory poisoning, intent breaking, cascading hallucinations, tool misuse, and agent communication poisoning. In the severity matrix, autonomous high-agency contexts make threats such as misaligned behaviors and unexpected remote code execution critical, each with a total score of $\sim 15\%$6. This suggests that greater agency expands not only functional capability but also the space of semantically misaligned yet physically plausible failures (Eslami et al., 18 Dec 2025).

Governance questions therefore become constitutive rather than peripheral. The systems-level AgV framework identifies safety, real-time control, public acceptance, ethical alignment, and regulatory frameworks as key challenges, and extends the discussion to accountability, cybersecurity, labor impacts, equity, and environmental effects. A common misconception is that more agency simply yields more intelligent mobility. The literature instead supports a stricter formulation: agency is valuable only when layered above deterministic safety gates, communication-aware control, explicit human override, and institutional mechanisms capable of governing vehicles that can reason, negotiate, and invoke tools (Yu, 7 Jul 2025).