Communication efficiency has emerged as a critical challenge in multi-agent systems (MAS), particularly as these systems scale to include hundreds or thousands of agents operating in bandwidth-constrained environments. Recent developments in 2024-2025 have introduced innovative solutions spanning communication compression, sparse protocols, information-theoretic methods, learned strategies, and edge computing architectures. These advances address the fundamental tension between the need for coordination and the practical limitations of wireless networks, making autonomous systems more scalable and robust for real-world deployment.
Modern multi-agent systems increasingly rely on sophisticated compression techniques to minimize data transmission while preserving critical information. EffiComm, a breakthrough framework accepted for ITSC 2025, demonstrates the potential of learned compression strategies for vehicle-to-everything (V2X) applications. The system employs a two-stage pipeline combining Selective Transmission (ST) to eliminate low-confidence regions with Adaptive Grid Reduction (AGR) using Graph Neural Networks to assign vehicle-specific compression ratios. On the OPV2V benchmark, EffiComm achieves 0.84 mAP@0.7 detection accuracy while transmitting only 1.5 MB per frame, representing less than 40% of the data required by previous methods.
Quantization has emerged as another powerful compression approach. The CACOM protocol, presented at AAMAS 2024, incorporates learned step size quantization (LSQ) to reduce bandwidth requirements while maintaining performance. More recently, QuantV2X introduces end-to-end quantization across entire collaborative perception pipelines, replacing costly FP32 bird's-eye-view (BEV) feature transmission with compact low-bit messages. Each agent transmits only code indices from a shared codebook, allowing receivers to reconstruct high-fidelity features locally.
Sparse communication protocols address bandwidth constraints by determining when and what information should be transmitted. The Model-Based Communication (MBC) framework, published in early 2025, uses supervised learning to build message estimation models that enable agents to decide whether communication is necessary based on predictability. Experimental results show MBC improves performance over state-of-the-art baselines while significantly reducing communication overhead.
Coordination graphs provide a natural framework for sparse communication in multi-agent reinforcement learning. The Influence Enhanced Sparse Coordination Graphs (IESCG) framework represents the first attempt to quantify collaborative values within coordination graph structures, constructing influence networks to depict interaction importance and building sparse coordination graphs accordingly.
The integration of world models offers an alternative approach to sparse communication. Research presented at NeurIPS 2025 demonstrates that communicating predicted plans rather than raw percepts achieves superior scalability. The proposed Intention Communication approach uses an Imagined Trajectory Generation Module (ITGM), a compact learned world model that simulates future states. In complex environments, this approach achieved 96.5% success compared to only 12.2% for learned direct communication, suggesting structured predictive models are essential for robust coordination at scale.
Information-theoretic approaches provide principled frameworks for learning minimal sufficient representations for communication. The MAGI (Multi-Agent communication via Graph Information bottleneck) method, published at AAAI 2024, optimally balances robustness and expressiveness by maximizing mutual information between message representations and selected actions while simultaneously constraining mutual information between message representations and agent features. This dual optimization ensures messages contain only task-relevant information, improving both communication efficiency and robustness to noise.
Building on these principles, the Consensus-Driven Event-Based Graph Information Bottleneck (CDE-GIB) method integrates communication graphs with information flow through GIB regularizers to extract more concise message representations. The framework incorporates a variable-threshold event-triggering mechanism to further minimize communication volume required for establishing consensus during interactions, achieving high computational efficiency by avoiding expensive inner-loop operations.
The shift from hand-designed to learned communication protocols represents a fundamental transformation in multi-agent systems. Attention-based mechanisms have proven particularly effective, enabling agents to learn when communication is necessary and how to integrate shared information for cooperative decision-making. The attentional communication model has shown remarkable results in large-scale systems, with implementations achieving 6.4% reduction in communication overhead while maintaining performance in sparse-reward environments.
Goal-oriented communication approaches align communication behavior with task-specific objectives. Research in 2024-2025 demonstrates that agents can learn to produce sparse message passing, event-triggered communication, or adaptive protocols through multi-agent reinforcement learning combined with sparsity and attention mechanisms. These learned strategies often outperform engineered protocols by adapting to specific task requirements and environmental conditions.
Edge computing architectures fundamentally reshape communication requirements in multi-agent systems by distributing computation closer to data sources. Collaborative inference approaches split deep neural network models between user equipment and edge servers, resulting in faster and more energy-efficient inference. Adaptive early-exit schemes achieve up to 24.6% latency reduction and 46.5% energy consumption reduction compared to state-of-the-art IoT ML inference by adaptively distributing computation between devices and edge servers.
Multi-agent systems benefit from agentic AI at the edge through techniques from multi-agent reinforcement learning, distributed constraint optimization, and graph neural network-based coordination. These approaches enable decentralized decision-making, allowing distributed agents to perform collaborative inference and task allocation without centralized control.
Robotic swarms present unique communication challenges due to scale, mobility, and resource constraints. Recent research demonstrates that purely vision-based approaches can achieve effective coordination without traditional communication infrastructure. A 2024 study presents decentralized terrestrial swarms where robots achieve polarized motion with collision avoidance exclusively through visual interactions, computing everything onboard based on individual camera streams. This approach eliminates vulnerability to communication jamming, delays, or outages while reducing overall system complexity.
When communication is employed, bio-inspired approaches show promising results. Hybrid communication architectures based on Robot Operating System (ROS) and Message Queuing Telemetry Transport (MQTT) protocols overcome communication constraints in swarm robotics. Performance improvements are substantial: swarm intelligence algorithms improved coordination efficiency by 35% in systems with 500 robots, while blockchain-based systems enhanced data-sharing efficiency by 42% in decentralized swarms supporting 200 robots with 98% data delivery success rates.
Industrial IoT applications face particularly severe bandwidth constraints due to high device density and limited computational resources. Multi-agent systems have become preferred architectures for IIoT due to their suitability for decentralized decision-making and enhanced security. Multi-agent deep reinforcement learning (MARL) addresses scenarios where numerous agents optimize long-term performance in partially observable environments, with each end device or edge node acting as an autonomous decision-making agent.
The period 2024-2025 has witnessed remarkable advances in communication efficiency for bandwidth-constrained multi-agent systems. Compression techniques like EffiComm and QuantV2X demonstrate that learned, adaptive strategies can reduce data transmission by 60% or more while maintaining performance. Information bottleneck methods provide principled frameworks for learning minimal sufficient representations, while sparse communication protocols and event-triggered mechanisms ensure transmission occurs only when necessary. The integration of edge computing and distributed inference architectures fundamentally reduces communication requirements by processing data closer to sources. Applications in V2X systems, robotic swarms, and industrial IoT validate these techniques in real-world scenarios, demonstrating scalability to hundreds or thousands of agents. As multi-agent systems continue growing in scale and complexity, these communication-efficient approaches will prove essential for practical deployment in bandwidth-constrained environments ranging from autonomous vehicles to smart cities.