multi-agent coordination

This work addresses the challenges of low coordination efficiency, lack of auditability, and insufficient policy compliance in multi-agent systems for complex tasks by proposing a unified orchestration framework that integrates planning, policy execution, state management, and observability mechanisms. The framework introduces two novel protocols—the Model Context Protocol and Agent2Agent Protocol—to enable standardized, extensible communication among agents and between agents and external tools. By supporting policy governance and end-to-end traceability, the proposed architecture provides a practical, transparent, and accountable blueprint for deploying large-scale multi-agent systems within enterprise AI ecosystems.

This work addresses the robust co-synthesis of joint action and communication policies for stochastic multi-agent systems under communication constraints, aiming to maximize the probability of reaching a common reach-avoid objective. We propose a novel information-overhead cost function, enabling— for the first time—the joint robust synthesis of action and communication policies. Our approach models the system as a stochastic game and integrates symbolic policy synthesis, constrained optimization, probabilistic reachability analysis, and quantitative information-flow metrics. The method rigorously guarantees performance bounds under dynamic bandwidth constraints and establishes both the existence and computational tractability of feasible policies. Evaluated on multiple benchmark tasks, the synthesized policies achieve over 92% of the unconstrained optimal reach-avoid probability while satisfying strict communication limits, thereby significantly advancing both the practical applicability and theoretical completeness of resource-constrained multi-agent coordination.

This work addresses the trade-off in multi-agent systems between performance degradation under uncoordinated execution and the high communication overhead of explicit coordination. The authors propose a novel on-demand coordination mechanism that conceptualizes coordination as a continuum ranging from joint optimization to Nash equilibrium. Crucially, they establish the first connection between coordination decisions and second-order properties of the objective function—specifically, the structure of the Hessian—by dynamically analyzing its curvature to determine when coordination is necessary. Integrating differentiable motion planning with game-theoretic principles, the proposed algorithm accurately identifies critical moments requiring coordination, thereby maintaining high task performance while substantially reducing communication costs.

To address stability and cooperative optimization challenges in decentralized control of large-scale distributed systems, this work integrates cooperative game theory with control theory to establish, for the first time, a systematic framework of coalition control. It embeds game-theoretic solution concepts—such as the Shapley value and core—directly into feedback control law design, thereby unifying performance allocation and strategic coordination. Methodologically, the approach combines distributed model predictive control (MPC), multi-agent modeling, and robust stability analysis to develop a cooperative decision-making mechanism under partial information sharing. Experimental validation on power grid and traffic flow control benchmarks demonstrates a 23% improvement in cooperative efficiency over conventional decentralized control, alongside significantly enhanced robustness against disturbances and uncertainties. The core contribution is a novel, game-theoretic-solution-driven distributed control paradigm with provable closed-loop stability guarantees.

To address low collaboration efficiency and poor adaptability in cross-subteam coordination within multi-agent systems, this paper proposes a bottom-up dynamic cooperative control framework. The method dynamically evolves communication-driven control coalitions in real time based on time-varying inter-subsystem coupling strength and embeds cooperative game mechanisms into a distributed model predictive control (MPC) architecture to jointly optimize coalition structure and controller parameters. Its key innovation lies in establishing, for the first time, a coupling-strength-guided coalition formation mechanism that simultaneously enhances coordination efficiency and robustness under competitive environments. Evaluation on canonical coupled systems demonstrates that the proposed approach significantly reduces overall control cost, improves response consistency and disturbance rejection capability, and adaptively adjusts communication overhead according to coupling degree—outperforming fixed-topology or static-coalition strategies.

Existing agent programming models struggle to coordinate internal decision-making with external interaction behaviors and lack effective abstractions for protocol adherence. This work proposes Kiko, a protocol-based agent programming model that introduces a decider abstraction, enabling agents to select actions and generate protocol-compliant messages exclusively from the set of legally permissible moves, thereby strictly adhering to interaction protocols in decentralized environments. Kiko fully encapsulates low-level communication details, supports flexible decision-making strategies, and integrates formal operational semantics to guarantee correct execution of arbitrary protocols. By doing so, it allows developers to focus on business logic while ensuring correctness and reliability in protocol enforcement within multi-agent systems.

Existing agent interoperability protocols primarily focus on task coordination and lack support for governance-constrained collective decision-making in multi-agent communities. Drawing on organizational theory and corporate governance standards, this work proposes a six-dimensional governance framework encompassing membership management, deliberation, voting, dissent reservation, human escalation, and audit replay. The study systematically evaluates five prominent protocols—MCP, A2A, ACP, and others—and reveals, for the first time, that the core deficiency lies not in insufficient protocol-level features but in the absence of a dedicated architectural layer for governance. It further distinguishes between scalability gaps and structural gaps. The analysis demonstrates that current protocols universally lack voting and dissent reservation mechanisms, provide only partial support for deliberation, and offer no complete set of governance primitives.

Existing approaches struggle to effectively model differentiated communication and coordination among heterogeneous agents, limiting overall team performance. This work proposes Heterogeneous Policy Network (HetNet), the first framework to enable end-to-end joint learning of collaborative policies and binary communication for large-scale heterogeneous robotic teams. Built upon a heterogeneous graph attention mechanism and integrated with multi-agent reinforcement learning, HetNet achieves substantial communication bandwidth reduction—up to 200-fold compression—while outperforming state-of-the-art baselines by 5.84% to 707.65% across multiple tasks, thereby simultaneously ensuring highly efficient communication and superior collaborative performance.

This work addresses the challenge of balancing structural stability and dynamic adaptability in large language model–driven multi-agent systems. The authors propose MACA, a novel framework that jointly models coordination structure and orchestration as a probabilistic inference problem. By leveraging a structure prior conditioned on task requirements and resource budgets, MACA enables fine-grained, controllable adaptive coordination through policy-guided orchestration. The approach maintains system stability while significantly enhancing efficiency, achieving an average performance gain of 8.42% across multiple benchmarks and reducing token consumption by 43.19%. This reduction effectively suppresses redundant agent interactions and accelerates convergence to high-efficiency execution pathways.

This study addresses the challenge of coordinating heterogeneous agents and selecting appropriate communication protocols in large language model–driven multi-agent systems. Drawing from software engineering practice, it empirically evaluates the coordination capabilities of the Model Context Protocol (MCP) and Agent2Agent (A2A) across critical scenarios including agent discovery, multi-turn dialogue, and asynchronous communication. The work reveals that MCP offers lightweight efficiency but requires explicit state management, making it suitable for low-complexity tasks, whereas A2A natively supports stateful coordination—better suited for strongly state-dependent contexts—at the cost of higher implementation complexity. These findings provide actionable empirical guidance for multi-agent protocol design and clarify key trade-offs between the two approaches regarding responsibility distribution, interoperability, and access control.