scalable deployment

This work addresses critical challenges in the real-world deployment of large language model (LLM)-driven agent systems, particularly concerning robustness, safety, and reliability. Bridging academic advances with industrial practice, the study presents case studies from software engineering, scientific discovery, and finance to distill reusable design patterns and an evaluation checklist. It integrates key techniques including LLM-based reasoning and planning, multi-agent coordination, validation pipelines, fallback mechanisms, and human-in-the-loop oversight. The proposed cross-domain deployment framework has been validated in pharmaceutical discovery and financial systems, demonstrating significant improvements in stability and trustworthiness of agent systems in real-world settings, thereby narrowing the gap between research innovation and practical implementation.

To address high deployment latency, external data transmission dependency, and poor adaptability of AI models in Industry 5.0 edge environments, this paper proposes AgentEdge—a lightweight, modular agent-based framework for rapid AI model deployment and real-time local inference across heterogeneous industrial edge devices. It employs a hierarchical agent architecture enabling human–machine collaborative task decomposition and dynamic resource scheduling, thereby significantly reducing end-to-end latency and eliminating the need to transmit sensitive data externally. With minimal resource overhead (<128 MB RAM), strong scenario adaptability, and plug-and-play integration capability, AgentEdge enhances the scalability and accessibility of edge AI. Evaluated on a real-world food manufacturing production line, it shortens model deployment time by 67% and reduces system response latency to under 83 ms. The open-source implementation demonstrates clear industrial applicability and practical deployment value.

To address the challenge of decentralized multi-agent coordination in cloud-edge heterogeneous environments lacking centralized server support, this paper proposes a decentralized multi-agent collaboration framework. The framework introduces a novel logistics-object modeling approach and an abstract agent interface, enabling dynamic service-flow orchestration, topology-aware distributed publish-subscribe communication, and many-to-many service election. It integrates plug-and-play node discovery, flexible task reconfiguration, and fault-adaptive agent replacement. By synergistically combining multi-agent systems (MAS), fault-tolerant scheduling, and dynamic topology management, the framework achieves autonomous, real-time, and scalable decision-making coordination without a central coordinator. Experimental results demonstrate significant improvements in system resilience, real-time performance, and robustness for mission-critical autonomous scenarios, confirming its capability for highly available deployment.

This study addresses the lack of systematic investigation into architectural design decisions for non-large language model components in current AI agent systems. The authors propose a protocol-guided, source code–driven empirical analysis method that enables, for the first time, transparent deconstruction of heterogeneous AI agent systems. Through cross-project qualitative coding and co-occurrence analysis of 70 open-source projects, they identify five core design dimensions—sub-agent architecture, context management, tooling systems, security mechanisms, and orchestration—and uncover their combinatorial patterns. Based on these findings, the study further distills five archetypal architectural patterns: lightweight tool-oriented, CLI framework–based, multi-agent orchestrator, enterprise system, and domain-specific vertical architectures.

This work addresses the current lack of open-source infrastructure capable of efficiently training and evaluating large-scale agents on complex tasks such as software engineering and computer operation. To this end, we propose a three-service decoupled architecture tailored for agent-environment interaction workloads, which separates the system into three independent services—model, agent, and environment—enabling fine-grained task scheduling, dynamic resource allocation, and unified interface communication. This design allows each component to scale independently and configure resources flexibly, significantly improving training efficiency and resource utilization. Experimental results demonstrate that the system can stably support tens of thousands of concurrent agent tasks, thereby filling a critical gap in infrastructure for large-scale agent training.

This study addresses the challenges of high latency, unstable concurrency, and security risks faced by large language model (LLM) agents in automating asset lifecycle management within Industry 4.0. The authors propose a Plan-then-Execute architecture that generates verifiable workflow graphs and integrates a topology-aware parallel scheduling mechanism to enable controlled inference overlap while ensuring functional correctness and security. Key technical contributions include topological-sort-based multi-agent scheduling, structured context pruning, dependency-aware concurrency control, and graceful degradation under fault injection. Evaluated on the AssetOpsBench benchmark, the system reduces median end-to-end latency by 1.6× (up to 1.8× for highly parallel tasks) and cuts inference overhead by approximately 30% through context pruning, all while maintaining stable task completion rates and output quality.

This work addresses the persistent challenge of inconsistent development and execution environments faced by researchers operating across heterogeneous computing platforms—ranging from laptops and workstations to supercomputers and cloud infrastructures. To overcome this, the authors propose a modular and portable software ecosystem featuring a unified command-line interface that enables seamless orchestration and execution of scientific workflows. The system ensures cross-platform consistency, reproducibility, and scalability, thereby streamlining computational research across diverse hardware configurations. Its practical efficacy has been demonstrated through successful integration into the plan4res project under the European Union’s Horizon 2020 initiative, where it effectively supported complex, large-scale scientific workflows in varied computing environments.

This work addresses the complexity and heavy reliance on manual effort in software architecture design, particularly the challenge of efficiently exploring diverse architectural solutions within agile development contexts. To this end, the paper proposes MAAD, a novel framework that systematically integrates multi-agent collaboration, Retrieval-Augmented Generation (RAG), and hierarchical memory mechanisms into the architectural design process. MAAD employs four specialized agents—analyst, modeler, designer, and evaluator—that collaboratively transform requirements into multi-view architectural blueprints and support iterative refinement. Experimental results across ten case studies demonstrate that MAAD-generated architectures consistently outperform those produced by the MetaGPT baseline in terms of completeness, modularity, and traceability. Furthermore, the evaluator agent automatically produces structured quality reports, substantially reducing the cost of manual validation.

This work addresses the challenges in edge and embedded application development—namely, heterogeneous software stacks, multi-language runtimes, and difficult debugging—which lead to rigid deployment workflows and complex fault diagnosis. To overcome these limitations, the paper proposes a novel architecture enabling unified end-edge-cloud development. Its core components include a single programming language, a retargetable runtime system, a local recording and replay mechanism for distributed events, and a cross-platform deployment framework. This design breaks down traditional debugging barriers in edge–cloud collaborative development, facilitating seamless scalability, consistent testing, and flexible deployment across heterogeneous environments. Evaluation of the prototype system demonstrates that the proposed approach significantly simplifies deployment procedures and enhances fault diagnosis efficiency.

This work addresses the challenges of dependency isolation, compatibility, reproducibility, and hardware resource sharing in multi-user collaborative and heterogeneous robotic deployments. To this end, it proposes a containerized architecture tailored for robot teams operating within edge–cloud协同 environments. The architecture uniquely integrates system-level containers (LXC/LXD), ROS 2/DDS communication middleware, and a three-tier edge infrastructure—comprising infrastructure core, platform orchestration, and compute acceleration—to enable topology-aware networking, strong isolation, and controllable resource sharing. Experimental validation in a real-world robotic laboratory demonstrates that the proposed approach significantly simplifies software integration, improves resource utilization, and supports secure prototyping alongside reproducible collaborative experimentation.