Wireless Multi-Agent Systems (WMAS)

• Wireless multi-agent systems (WMAS) are defined as networks where multiple agents interact via wireless links to jointly optimize communication, control, and inference.
• They encompass diverse architectures such as QoS-driven routing, TDMA broadcasting, and DAG-based AI orchestration to manage network functions.
• Practical implementations demonstrate robust connectivity, dynamic channel access, and enhanced performance using techniques like reinforcement learning and game-theoretic optimization.

Wireless multi-agent system (WMAS) denotes, in the recent literature, a class of systems in which multiple agents interact through wireless links, provide wireless infrastructure, or jointly optimize wireless-network functions. The term covers at least three closely related formulations: multi-agent intelligent wireless networks organized through game-theoretic learning (Wang et al., 2018), mobile wireless infrastructure on demand in which task agents and network agents jointly sustain end-to-end quality of service (QoS) (Calvo-Fullana et al., 2023), and AI-native wireless networks in which a meta agent or multiple specialized agents provide intelligent and customized services for different user equipment and optimize their conversation topology as a directed acyclic graph (DAG) (Peng et al., 1 Aug 2025). This suggests that WMAS is best understood not as a single protocol stack but as a design space spanning coordination, communication, sensing, control, and inference over wireless media.

1. Definition and conceptual scope

A useful organizing principle is the distinction between agents for communications and communications for agents. In the first category, agents participate in communication-system design and operation, including agent-generated communication software and agent-driven adaptive wireless optimization. In the second, wireless service supports agent operation, including network-supported single-agent loops and network-assisted multi-agent coordination (Liu et al., 15 May 2026). WMAS therefore includes both systems in which wireless networking is the object of optimization and systems in which wireless networking is the substrate that permits coordinated agency.

Task-oriented communication sharpens this view by replacing “send everything reliably” with “send only what the task needs, when it needs it.” In that formulation, timeliness is captured directly through the Age of Information, $\mathrm{AoI}_i(t)=t-\tau_i(t)$, rather than by raw latency alone, and semantic relevance becomes part of the control problem (He, 2022). This widens the scope of WMAS beyond conventional bit-pipe abstractions.

Another strand treats WMAS as an AI-native orchestration layer for intelligent and customized wireless networks. There, a meta agent selects the number of agents $K$, their roles, the number of rounds $T$, and an initial conversation topology $\mathcal{C}$, while a customized multi-agent system executes the request under that topology (Peng et al., 1 Aug 2025). A plausible implication is that WMAS has become a meeting point for wireless systems, multi-agent control, and agentic AI.

2. Architectural patterns

Representative WMAS architectures are heterogeneous in agent roles, control loci, and communication semantics.

Architecture	Agents or components	Control or interaction pattern
Mobile wireless infrastructure on demand (Calvo-Fullana et al., 2023)	Task agents; network agents / MID UAVs; designated planning agent	QoS-driven routing and waypoint replanning
BLAS (Shi et al., 2019)	Parent agents; child agents	Broadcast, passive receiving, D-TDMA
RadioMaster (Lei et al., 1 Jun 2026)	RadioWiki; RadioAgent; RadioEmulator; Operator	Retrieval-grounded synthesis and closed-loop verification
Agent-native wireless communications (Liu et al., 15 May 2026)	Device, edge, cloud; SMO; rApp; xApp	Agents for communications / communications for agents
WMAS for customized wireless networks (Peng et al., 1 Aug 2025)	Meta agent; customized MAS	DAG-based conversation topology

In mobile wireless infrastructure on demand, task agents are end-users such as UAVs on a mission, ground robots, livestock tags, or humans with hand-held radios, while network agents are autonomous quadrotors whose physical positions $x_j\in\mathbb{R}^3$ and routing tables are under system control. A designated planning agent periodically gathers positions via GPS and ROS multi-master and, without any centralized base station, computes optimal probabilistic routing and next-waypoint positions (Calvo-Fullana et al., 2023).

BLAS adopts a different asymmetry. Parent agents broadcast UWB packets and child agents are passive receivers. The asynchronous broadcasting and passively receiving protocol schedules parent broadcasts with distributed TDMA, and both parent and child agents run distributed state-estimation procedures for joint relative localization and clock synchronization (Shi et al., 2019).

RadioMaster exemplifies a WMAS in which the output is not only network behavior but physical radio emission. RadioWiki supplies multimodal domain knowledge, RadioAgent decomposes tasks and produces complex I/Q samples, RadioEmulator verifies the waveform in a virtual transmitter-channel-receiver chain, and the Operator deploys only validated waveforms to SDR hardware (Lei et al., 1 Jun 2026).

3. Mathematical models and optimization structure

One canonical WMAS model defines a graph $V=\{1,\dots,L\}=V_T\cup V_N$, with task agents $V_T$ and network agents $V_N$. Routing variables $\alpha_{ij}^k\in[0,1]$ specify the fraction of time, or probability, that agent $i$ forwards flow $K$0 to agent $K$1. Each flow demands a QoS pair $K$2, and the outage constraint is $K$3. With fixed positions $K$4, robust routing is posed as a convex Second-Order Cone Program; with fixed $K$5, connectivity maximization is posed as an SDP over the Fiedler value $K$6 of the graph Laplacian (Calvo-Fullana et al., 2023).

In AI-native WMAS, the conversation graph is explicitly modeled as a DAG $K$7 with $K$8 and adjacency matrix $K$9. The optimization criterion is $T$0, and a heatmap-based REINFORCE procedure optimizes edge probabilities while enforcing structural constraints such as no self-loops, no backward edges across rounds, and no edges beyond the next round (Peng et al., 1 Aug 2025).

Game-theoretic learning work places this modeling effort across pattern recognition, prediction, and decision making, and distinguishes internal coordination for resource optimization from external adversarial decision-making for anti-jamming under constrained capabilities, incomplete information, and heterogeneous wireless networks (Wang et al., 2018). Deep reinforcement learning extends the same agenda by casting WMAS problems as MDPs or Dec-POMDPs under CTDE, federated, or DTDE regimes, while wireless-sensor-network studies survey minimax, Nash, Stackelberg, and mean-field variants inside the Q-update for resource allocation and task scheduling (He, 2022, Wu et al., 2024, Tashakori, 2021).

4. Wireless coordination mechanisms

A distinctive WMAS theme is the use of wireless-channel physics as part of the algorithm. For average consensus over a wireless multiple-access channel, each agent broadcasts two orthogonal real symbols, $T$1 and $T$2, receives

$T$3

and $T$4, and updates $T$5. Under a strongly connected graph, consensus is always reached, although the consensus value depends on channel variations (Molinari et al., 2018).

For max-consensus, the superposition property is combined with an authorization bit $T$6. Agents broadcast $T$7 and $T$8, compute the authorized-neighbor average $T$9, and update $\mathcal{C}$0. A periodic dyadic reset of the authorization state yields finite-time max-consensus in the connected undirected case (Molinari et al., 2018).

Other WMAS protocols use broadcast and wireless impairments directly. BLAS uses one-way TOA measurements, pseudo-clock-offset Kalman filtering, and iterative least-squares localization within its ABPR/D-TDMA broadcast structure (Shi et al., 2019). Diffusion adaptation over wireless links inserts equalization coefficients $\mathcal{C}$1 into the diffusion combination step, updates combination weights dynamically as neighborhoods change under fading, and uses pilot-aided channel coefficient estimates when CSI is unavailable (Abdolee et al., 2015). These designs show that interference, fading, path loss, and asynchronous broadcast are not treated only as impairments; they are also algorithmic primitives.

5. Application domains and empirical evidence

On-demand mobile wireless infrastructure is one of the clearest WMAS embodiments. In simulation, a U-shaped trajectory over $\mathcal{C}$2 produced range extension from $\mathcal{C}$3; in a cloverleaf patrol, a fixed setup suffered $\mathcal{C}$4 outages while MID achieved $\mathcal{C}$5 outage; and after one network UAV was deactivated and swapped, MID maintained connectivity with no outage while a static deployment would suffer a $\mathcal{C}$6 outage. In real-world UAV experiments, MID supported $\mathcal{C}$7 to $\mathcal{C}$8 while static support extended only to $\mathcal{C}$9; during a circular patrol of diameter $x_j\in\mathbb{R}^3$0, throughput stabilized at $x_j\in\mathbb{R}^3$1 with delay $x_j\in\mathbb{R}^3$2 after $x_j\in\mathbb{R}^3$3; and relay replacement recovered average rate from $x_j\in\mathbb{R}^3$4 without human intervention (Calvo-Fullana et al., 2023).

BLAS illustrates a different empirical regime. In simulation with 5 parents and 3 children, parent-position RMSE was 0.9–2.7 cm and child-position RMSE was ∼10–11 cm. In hardware, a parent test reported range RMSE = 0.051 m and clock-sync ≈0.17 ns, while a child test reported inter-child distance RMSE = 0.064 m (Shi et al., 2019).

RadioMaster extends WMAS from network control to autonomous radio signal generation. RadioBench reports Level 1 overall QAA of 0.94, Level 2+3 overall CPR of 0.83 versus a best baseline of 0.35, HDR of 0.81 versus ∼0.32, and SIR of 0.71 versus near-zero for all baselines. Its architecture couples retrieval-grounded protocol and hardware knowledge, collaborative I/Q sample generation, and pre-deployment physical-layer verification (Lei et al., 1 Jun 2026).

Resource-management applications remain central. Multi-agent DRL for joint dynamic channel access and power control reaches 90% of the performance achieved by the combination of weighted minimum mean square error algorithm for power control and exhaustive search for dynamic channel access (Lu et al., 2021). Multi-agent reinforcement learning for wireless network protocol synthesis learns to adjust MAC layer transmission probabilities, attains theoretical maximum throughput at an optimal load, and retains that maximum throughput at higher loading conditions while remaining agnostic to heterogeneous loading and allowing parametrically adjusted access priorities (Dutta et al., 2021). In MU-MIMO-OFDMA uplink scheduled access, MAxLM reports that Prompt Template 1 yields <1% assignment error, whereas Template 2 yields approximately 40–45%, and Mistral-NeMo:12b achieves up to 30% UL sum-rate gain over the best-channel-quality heuristic (Quadri et al., 15 May 2026).

6. Limitations, misconceptions, and open problems

No single deployment model exhausts WMAS. Mobile Wireless Infrastructure on Demand explicitly assumes the responsibility of creating and sustaining a wireless network capable of satisfying end-to-end communication requirements of a task team, yet its framework lists centralized architecture as a limitation and notes that sampling-based repositioning may not scale gracefully for very large $x_j\in\mathbb{R}^3$5 or fast dynamics (Mox et al., 2020). The literature also makes clear that WMAS is not reducible to throughput maximization: evaluated quantities include outage, algebraic connectivity, BER, packet-error or frame-error rate, traceroute delay, AoI, task-success rate, average per-packet delay, packet-loss rate, energy consumption, and conversation overhead (Lei et al., 1 Jun 2026, He, 2022, Meng et al., 27 Oct 2025, Peng et al., 1 Aug 2025).

Several research agendas recur. Game-theoretic learning papers emphasize constrained capabilities of agents, incomplete information obtained from the environment, and the distributed, dynamically scalable and heterogeneous characteristics of wireless network, with anti-jamming and adversarial decision-making as continuing concerns (Wang et al., 2018). The mobile-infrastructure line identifies fully decentralized implementations of both SOCP and SDP, non-Gaussian channel models, joint optimization of utility and connectivity, lower-level flight-dynamics constraints, and security and jamming resilience as open questions (Calvo-Fullana et al., 2023). Agent-native wireless communications adds scalability and conflict-aware control, mobility-aware loop continuity, loop-centric KPIs, standardized northbound agent APIs, and token-/task-oriented communications (Liu et al., 15 May 2026). Work on wireless multi-agent generative AI adds high inference latency, privacy and security, unreliable coverage, and massive backhaul overhead as limitations of cloud-based LLMs, motivating on-device LLMs and domain-specialized TelecomLLMs (Zou et al., 2023).

This suggests that WMAS is converging toward systems in which wireless communication, agent coordination, and domain-specific decision loops are co-designed rather than layered independently. In that form, WMAS is not merely a wireless extension of multi-agent systems, nor merely an AI layer atop communications; it is a joint framework for networked agency under wireless constraints.