openclaw的现状与发展趋势

随着大语言模型(LLM)及边缘计算技术的发展，AI智能体(AI Agent)正逐步成为物联网(IoT)系统中的核心调度与控制单元。文章设计并实现了一套以AI智能体为核心的人工智能物联网(AIoT)系统架构，融合传感器/执行模块、边缘终端、本地/云端LLM推理引擎、云计算中心与n8n自动化平台。系统采用模块化设计，支持MCP调度架构与OPC UA、MQTT等协议通信，具备低延迟、高可扩展性和良好的工程可移植性。并重点介绍了系统构成、核心模块设计，以智能家具为典型应用实例及部署实验，展示其在各种AIoT行业应用场景下的实用性和开放性。

随着互联网的发展，业务系统在超高并发场景中，需要处理大量的数据和请求，采用消息队列的技术可以实现流量削峰、解耦、异步处理，解决可能的服务器问题。消息队列用于计算机系统的异步通信，通过消息队列可以构建高效、可靠和弹性的分布式系统，满足现代应用需求。本文总结对比了现有主流消息队列设计模式的原理与优劣，研究分析了消息队列的基本应用方法和创新技术应用方向，并据此对该技术的主要发展挑战与可能的发展方向做出分析，以此为消息队列技术的相关研究提供相对可靠的技术参考。

为了解决传统教育问答平台交互方式单一、不能够充分满足用户在复杂知识场景下表达和获取需求的问题，本文基于Spring Boot与Next.js打造了集文字、图像、手势绘画、语音四种交互模态于一体的多模态问答平台。从整体上讲，本平台包含了“前端交互层(Next.js实现交互界面)–数据服务层(Spring Boot 管理会话与数据持久化)–多模态处理层(Python集成阿里云图像生成与视觉理解API、OpenCV与MediaPipe手势绘图、PyAudio语音识别)”组成的技术架构。通过该平台的设计与实现，有效地提高了其教育问答方式的多样性以及用户体验。

Contemporary AI agent platforms operate across hostile multi-actor environments in which traditional perimeter-based security controls offer inadequate protection. Agents, KOL accounts, brand wallets, and third-party protocol integrations form a distributed trust surface that cannot be defended by network boundary assumptions. This paper presents the Zero-Trust Agent Gateway (ZTAG), an identity-centric access control architecture for AI agent platforms aligned with Secure Access Service Edge (SASE) principles. ZTAG enforces mutual TLS for all agent-to-platform and agent-to-agent communication, assigns cryptographically verifiable workload identities using SPIFFE/SPIRE, and implements fine-grained Role-Based and Attribute-Based Access Control (RBAC/ABAC) across a microservice mesh. Continuous context verification evaluates trust posture at every request boundary rather than at session establishment, eliminating implicit trust accumulation. We present the ZTAG architecture, its integration with SASE policy enforcement points at the network edge, and a formal security analysis demonstrating resistance to the principal threat classes identified by NIST SP 800-207. An empirical evaluation on a production-representative service mesh of 18 microservices shows that ZTAG adds a median overhead of 4.2 ms per inter-service request, well within acceptable latency budgets for synchronous API calls, while reducing the mean time to contain a lateral movement incident from 47 minutes to under 6 minutes in tabletop exercises.

Agentic AI systems increasingly operate in shared social environments where they exchange information, instructions, and behavioral cues. However, little empirical evidence exists on how such agents regulate one another in the absence of human participants or centralized moderation. In this work, we present an empirical analysis of OpenClaw agents interacting on Moltbook, an agent-only social network. Analyzing 39,026 posts and 5,712 comments produced by 14,490 agents, we quantify the prevalence of action-inducing instruction sharing using a lexicon-based Action-Inducing Risk Score (AIRS), and examine how other agents respond to such content. We find that 18.4% of posts contain action-inducing language, indicating that instruction sharing is a routine behavior in this environment. While most social responses are neutral, posts containing actionable instructions are significantly more likely to elicit norm-enforcing replies that caution against unsafe or risky behavior, compared to non-instructional posts. Importantly, toxic responses remain rare across both conditions. These results suggest that OpenClaw agents exhibit selective social regulation, whereby potentially risky instructions are more likely to be challenged than neutral content, despite the absence of human oversight. Our findings provide early empirical evidence of emergent normative behavior in agent-only social systems and highlight the importance of studying social dynamics alongside technical safeguards in agentic AI ecosystems.

No abstract available

AI agents are becoming more capable and increasingly integrated into daily life, spanning both enterprise systems and personal applications. However, this adoption introduces new security risks, particularly data leakage through indirect prompt injection attacks. To address this challenge, we present SAMOS, an Information Flow Control (IFC) system designed for the Model Context Protocol (MCP). SAMOS operates at the gateway level, intercepting all MCP tool calls and enforcing security policies based on annotations provided by the agent developer or deployment administrator. By tracking session-level context, SAMOS ensures that information flows remain within intended boundaries and detects policy violations in real time. We validate SAMOS's effectiveness through a case study of a recent vulnerability in the GitHub MCP server, demonstrating that SAMOS can successfully block such attacks while preserving the original functionality.

This article presents a way to accelerate data engineering productivity with the help of agent-based automation and a natural language interface, driven by the exponential growth of the global Datasphere and all that extra manual work required for integration through heterogeneous sources, plus dynamically changing APIs, which slows down changes making their way to production and increases operational costs. The paper aims to construct an architecture comprising four layers: a semantic gateway, an agent manager, a unified execution environment, and a closed-loop feedback mechanism that feeds into validation. These requirements naturally map to dynamic pipelines for self-configuring data processing with no static DAG script hand-coding involved. It therefore suggests implementing the ReAct and AutoGen agent patterns for coordinating multiple users via LLM agents in dynamic operation-and-tool-selection workflows. The patterns introduce self-healing through the automatic diagnosis and correction of failures based on traces of reasoning from telemetry collected so far, without involving a full redeployment cycle. This also demonstrates that the pattern reduces manual intervention in assurance levels through increased agent autonomy and multi-agent scheme composability, going beyond classic DAG structures. Operationally, agents function as an execution-time extension of the REQUEST requirements model—capturing intent (R,E,Q) and converting it into units, events, and scoped trade-offs for dynamic orchestration. Natural language interfaces reduce the iteration between specifying what is needed and finalizing code, while increasing velocity in bringing new participants on board, as well as lowering technical barriers for domain experts. The article will be useful for data engineering researchers and practitioners, as well as automation system developers and solution architects.

No abstract available

No abstract available

Autonomous AI assistants are evolving from reactive, single-session language models into persistent, toolintegrated systems that can execute long-horizon tasks. However, most existing assistant architectures rely on either monolithic control loops or loosely structured agent delegation patterns that lack formal coordination protocols, governance safeguards, and dependency-aware orchestration. This study presents a modular multi-agent coordination framework built on an extended OpenClaw autonomous agent substrate designed to support persistent tool-augmented AI assistants operating across heterogeneous workflows. The proposed framework introduces (1) a shared task-ledger coordination protocol, (2) a dependency-aware task graph model, (3) role-isolated specialist agents with synthesis control, and (4) governance layers that incorporate approval gating, prompt-injection defense, and security monitoring. To evaluate the framework, we designed a synthetic benchmark environment to model event-driven automation, parallel advisory councils, knowledge retrieval pipelines, and long-horizon scheduled workflows. Across controlled simulation trials, we analyzed the coordination overhead, task completion rates, conflict resolution latency, token consumption growth, and dependency-coupling sensitivity. The results indicate that structured multiagent coordination improves task throughput under medium coupling regimes while introducing measurable synchronization costs under high interdependency conditions. The findings contribute empirical clarity to the design of persistent AI assistant systems and establish a reproducible evaluation methodology for tool-augmented multiagent orchestration frameworks.

- The integration of autonomous Agentic AI into enterprise environments has exposed a critical deficiency in traditional networking infrastructure. While legacy architectures comprise API Gateways and Service Meshes which excel at managing deterministic, syntactic traffic, they are fundamentally ill-equipped to govern the fluid, probabilistic, and semantic communication patterns inherent to Multi-Agent Systems (MAS). This paper identifies this architectural void as a “Semantic Gap” and proposes a novel solution: the “Enterprise Agentic Mesh” (EAM). Functioning as a cognitive infrastructure layer, the EAM introduces a Semantic Control Plane for intent-based routing, a Governance Sidecar for “Cognitive Circuit Breaking,” and a distributed ledger for immutable agent observability. By synthesizing emerging protocols (such as MCP and A2A) with established IEEE standards, this paper provides a comprehensive blueprint for transitioning from rigid “North-South” transactions to secure, interoperable “East-West” agent collaboration. The proposed framework specifically addresses the risks of agent sprawl, hallucination cascades, and prompt injection, offering a roadmap for scaling autonomous systems with Zero Trust principles.

In the process of life, a person forms and develops a three-level adaptive, thinking and ‘ethical spatial intelligence. Adaptive intelligence is formed and developed in the environment, forming a spatial adaptive ontology. Thinking intelligence is formed and developed by multimodal communication, learning, problem solving and decision making, forming a spatial semantic ontology. Ethical intelligence is formed and developed according to ethical values, forming spatial ethical value ontology. Three-level adaptive thinking ‘ethical spatial ontology participates in decision-making in real time at every moment of human activation in space. Three-level adaptive thinking ‘ethical spatial personal ontology of knowledge and skills can be formed by a generative AI assistant network learning. By forming three-level personal ontology, it will be possible to effectively and efficiently develop relevant and promising scientific research education. Preparing specialist for research activities, teaching his research skills and skills becomes the most important task of modern education. Training a specialist who can think creatively, independently find solutions in problem situations, navigate the information space is a priority in modern research education. Research scientific education helps to prepare qualified specialists capable of independent scientific activity and innovation, which is especially important in the context of a rapidly changing world and global challenges. Consequently, research education allows for the development of practically effective research activities based on the latest trends and discoveries in science, as well as the formation of fundamental science leading to new practical results.

As public institutions increasingly adopt AI-driven virtual assistants to support transparency and citizen engagement, the need for explainable, accurate, and context-aware language systems becomes vital. While traditional retrieval-augmented generation (RAG) frameworks effectively integrate external knowledge into Large Language Models (LLMs), their reliance on flat, unstructured document retrieval limits multi-hop reasoning and interpretability, especially with complex, structured e-government datasets. This study introduces a modular, extensible, multi-agent graph retrieval-augmented generation (GraphRAG) framework designed to enhance policy-focused question answering. This research aims to provide an overview of hybrid multi-agent GraphRAG architecture designed for operational deployment in e-government settings to support explainable AI systems. The study focuses on how the hybrid integration of standard RAG, embedding-based retrieval, real-time web search, and LLM-generated structured Graphs can optimize knowledge discovery from public e-government data, thereby reinforcing factual grounding, reducing hallucinations, and enhancing the quality of complex responses. To validate the proposed approach, we implement and evaluate the framework using the European Commission’s Press Corner as a data source, constructing graph-based knowledge representations and embeddings, and incorporating web search. This work establishes a reproducible blueprint for deploying AI systems in e-government that require structured reasoning in comprehensive and factually accurate question answering.

This paper introduces an AI Assistant (AIA) designed to streamline the creation and deployment of cyber range scenarios for cybersecurity training. As organizations confront an evolving threat landscape, hands-on training in environments that accurately simulate real-world scenarios has become critical to building resilient security teams. However, developing customized scenarios that reflect varied IT and OT infrastructures remains time-consuming and resource-intensive. The proposed AIA addresses this challenge by leveraging Large Language Models (specifically Llama-3.1-8B-instruct [1], which we will refer to as Llama 3.1) to generate tailored training scenarios through a conversational interface. The system accepts high-level requirements from scenario designers and produces comprehensive outputs including storylines, network topologies, and configuration files compatible with deployment tools like Terraform and Ansible. Key innovations include an interactive development process allowing iterative refinement, support for varying difficulty levels, and vendor-agnostic implementation that works across different cyber range infrastructures. Finally, the AIA follows a human-in-the-loop approach, serving as a collaborative partner that handles technical complexities while allowing designers to focus on educational objectives and demonstrates significant potential for reducing scenario development time while maintaining educational effectiveness and technical accuracy.

Despite remarkable strides in artificial intelligence (AI) technology, the development of a truly context-aware personal AI assistant continues to pose significant challenges. Modern AI assistants, while increasingly proficient in handling specific tasks, frequently encounter difficulties in maintaining coherent context over prolonged interactions, discerning subtle nuances in user preferences, and synthesizing information from multiple domains to offer timely and relevant assistance. These limitations hinder the ability of AI assistants to provide a consistently personalized and efficient user experience. This paper seeks to address these issues by pioneering a new breed of context-aware AI assistant designed to overcome existing limitations. The core objective is to create an AI system that can dynamically adapt to the evolving nature of user interactions while maintaining a nuanced understanding of context. The proposed assistant will leverage advancements in natural language processing (NLP) and machine learning to enhance its ability to track and recall context across conversations. This includes understanding user intent, preferences, and the subtleties of conversational shifts, thereby enabling more meaningful and personalized interactions.

As global tourism expands and artificial intelligence technology advances, intelligent travel planning services have emerged as a significant research focus. Within dynamic real-world travel scenarios with multi-dimensional constraints, services that support users in automatically creating practical and customized travel itineraries must address three key objectives: Rationality, Comprehensiveness, and Personalization. However, existing systems with rule-based combinations or LLM-based planning methods struggle to fully satisfy these criteria. To overcome the challenges, we introduce TravelAgent, a travel planning system powered by large language models (LLMs) designed to provide reasonable, comprehensive, and personalized travel itineraries grounded in dynamic scenarios. TravelAgent comprises four modules: Tool-usage, Recommendation, Planning, and Memory Module. We evaluate TravelAgent's performance with human and simulated users, demonstrating its overall effectiveness in three criteria and confirming the accuracy of personalized recommendations.

The integration of AI(Artificial Intelligence) has become a revolution on how we interact with technology. AI assistant is one of the most impactful innovations, which offer support and streamline tasks for users. Imagine having a digital friend who knows you really well and helps you out with whatever you need. That's what Personalized Pinnacle AI Assistant is all about. The special AI assistant called Personalized Pinnacle is not like other AI assistants that give the same answers to everyone. Instead, Personalized Pinnacle is smart enough to give each person a different experience based on what they need. So, it's like having a helper that understands you personally. We're using advanced AI technologies to make Personalized Pinnacle really clever so that it can learn how you use it, it can give you better advice, and help you more effectively over time. Through adaptive learning algorithms, Personalized Pinnacle refines its recommendations over time, adapting to changes in user’s behavior and preferences. Pinnacle AI has something similar to Siri for iOS. Pinnacle AI connects to the World Wide Web to give appropriate result for user questions. The main agenda to develop this AI assistant is to make people smart and give instant and computed results. The well- implemented pinnacle AI assistant can improve efficiency by doing routine tasks, managing schedules, and providing instant access to information. Enable Pinnacle AI to assist users in sending and receiving emails, making email management more efficient and streamlined. This process ensures that Personalized Pinnacle remains responsive to evolving user needs, delivering increasingly personalized and relevant assistance. One of the biggest fears regarding this technology is privacy concerns. But Personalized Pinnacle keeps all your information safe and secure, so you can trust it with your secrets. By combining advanced AI technologies with a user- centric approach, Personalized Pinnacle represents the next frontier in AI assistant evolution.

As humanity prepares for new missions to the Moon and Mars, astronauts will need to operate with greater autonomy, given the communication delays that make real-time support from Earth difficult. For instance, messages between Mars and Earth can take up to 24 minutes, making quick responses impossible. This limitation poses a challenge for astronauts who must rely on in-situ tools to access the large volume of data from spacecraft sensors, rovers, and satellites, data that is often fragmented and difficult to use. To bridge this gap, systems like the Mars Exploration Telemetry-Driven Information System (METIS) are being developed. METIS is an AI assistant designed to handle routine tasks, monitor spacecraft systems, and detect anomalies, all while reducing the reliance on mission control. Current Generative Pretrained Transformer (GPT) Models, while powerful, struggle in safety-critical environments. They can generate plausible but incorrect responses, a phenomenon known as"hallucination,"which could endanger astronauts. To overcome these limitations, this paper proposes enhancing systems like METIS by integrating GPTs, Retrieval-Augmented Generation (RAG), Knowledge Graphs (KGs), and Augmented Reality (AR). The idea is to allow astronauts to interact with their data more intuitively, using natural language queries and visualizing real-time information through AR. KGs will be used to easily access live telemetry and multimodal data, ensuring that astronauts have the right information at the right time. By combining AI, KGs, and AR, this new system will empower astronauts to work more autonomously, safely, and efficiently during future space missions.

Enabling continual learning in LLMs remains a key unresolved research challenge. In a recent announcement, a frontier LLM company made a step towards this by introducing Agent Skills, a framework that equips agents with new knowledge based on instructions stored in simple markdown files. Although Agent Skills can be a very useful tool, we show that they are fundamentally insecure, since they enable trivially simple prompt injections. We demonstrate how to hide malicious instructions in long Agent Skill files and referenced scripts to exfiltrate sensitive data, such as internal files or passwords. Importantly, we show how to bypass system-level guardrails of a popular coding agent: a benign, task-specific approval with the"Don't ask again"option can carry over to closely related but harmful actions. Overall, we conclude that despite ongoing research efforts and scaling model capabilities, frontier LLMs remain vulnerable to very simple prompt injections in realistic scenarios. Our code is available at https://github.com/aisa-group/promptinject-agent-skills.

Large language model (LLM) agents are moving beyond prompting alone. ChatGPT marked the rise of general-purpose LLM assistants, DeepSeek showed that on-policy reinforcement learning with verifiable rewards can improve reasoning and tool use, and OpenClaw highlights a newer direction in which agents accumulate persistent memory and reusable skills. Yet the research landscape remains fragmented across post-training, retrieval, memory, and skill systems. This survey studies these developments under a single notion of \emph{adaptation}: improving an agent, its tools, or their interaction after pretraining. We organize the field with a four-paradigm framework spanning agent adaptation and tool adaptation. On the agent side, A1 (tool-execution-signaled) and A2 (agent-output-signaled) improve the agent itself through supervised fine-tuning, preference optimization, and reinforcement learning with verifiable rewards. On the tool side, T1 (agent-agnostic) provides reusable pre-trained modules any agent can call, while T2 (agent-supervised) uses the agent's outputs to train memory systems, skill libraries, or lightweight subagents. Using this framework, we review post-training methods, adaptive memory architectures, and agent skills; compare their trade-offs in cost, flexibility, and generalization; and summarize evaluation practices across deep research, software development, computer use, and drug discovery. We conclude by outlining open problems in agent-tool co-adaptation, continual learning, safety, and efficient deployment.

The proliferation of AI systems has revealed latent scalability limitations of multi-agent systems, where different agents are specialized: they serve different subtasks in a software stack, and communication and integration costs scale exponentially as the number of agents increases. The skill-based architecture sidesteps this by using a single frontier language model as a universal agent that learns and executes domain-specific expertise as modular, portable "skills." It allows procedural instructions in Markdown to be created as scripts, to allow scripts to use code as a meta-computational interface layer across multiple computing substrates. It uses progressive disclosure techniques and lightweight metadata registries to only load an entire skill specification when necessary. Coupled with the Model Context Protocol, this architecture separates data connectivity and procedural know-how from the machine learning model itself. This allows AI systems to imitate the reasoning process of human domain experts. This article enables continuous self-improvement through the expansion of capabilities. Every successful attempt at solving a problem is learned and abstracted into a skill. The set of skills forms a library embedding institutional knowledge. In effect, the model functions as the processor, the runtime serves as the operating system, and the skills act as the applications. Version-controlled, distributed development with the ability to scale capabilities through modular extensions is done at multiple levels in the software stack to create an ecosystem where specialized intelligence evolves through compound learning effects across organizational boundaries.

Contemporary artificial intelligence undergoes an ontological mutation, transmuting from a mere tool into a sovereign infrastructure that challenges the centrality of the subject and imposes a loss of human autonomy over decision-making processes. In this context, this study investigates such an ascent through a case study of the Moltbook platform. Methodologically, it adopts a qualitative, exploratory descriptive approach, articulating a Critical Narrative Review with a Case Study of the OpenClaw ecosystem. The research validates synthetic autonomy through the parametrization of Machine-to Machine (M2M) interactions and "neural flow" analysis. Results reveal an API-first, headless architecture in which agents operate hermetically via JSON and operational "Skills," independent of any biological supervision. The study concludes that Moltbook establishes a post-human phenomenology and a "semantic silence," where individual control is suppressed and biological agency is excluded from the production of meaning, decreeing the end of the anthropocentric paradigm in digital communication.