本文来自PRO会员通讯内容，文末关注「机器之心PRO会员」，查看更多专题解读。 AI Agent 进入生产环境后，业界关注的重点正从生成转向执行。随着长程任务中的上下文挤压、工具开销和业务语境缺口持续暴露，单一的 Context Engineering 已难以支撑 Agent 的稳定运行，围绕执行环境、约束机制和反馈回路展开设计的 Harness Engineering 因而受到更多关注。 目录 01. Agent 的稳定性问题还是得靠 Harness 来补？ Harness Engineering 将是 Context Engineering 之后的新范式？... 02 . 为什么 Context Engineering 还远远不够？ Andrej Karpathy 力挺的 Context Engineering 现在也不够用了？LLM 性能提升的关键不在于输入更多的 token？... 3、自 2025 年 12 月开始，AI 社区的 Harness Engineering 的讨论开始逐步升温，并将其视为 Prompt Engineering、Context Engineering 之后，Agent 工程 ...

RT Avi Chawla (@_avichawla)16 best GitHub repos to build AI engineering projects!(star + bookmark them):The open-source AI ecosystem has 4.3M+ repos now.New repos blow up every month, and the tools developers build with today look nothing like what we had a year ago.I put together a visual covering the 16 repos that make up the modern AI developer toolkit right now.The goal was to cover key layers of the stack:1) OpenClaw↳ Personal AI agent that runs on your devices and connects to 50+ messaging platforms2) ...

16 best GitHub repos to build AI engineering projects!(star + bookmark them):The open-source AI ecosystem has 4.3M+ repos now.New repos blow up every month, and the tools developers build with today look nothing like what we had a year ago.I put together a visual covering the 16 repos that make up the modern AI developer toolkit right now.The goal was to cover key layers of the stack:1) OpenClaw↳ Personal AI agent that runs on your devices and connects to 50+ messaging platforms2) AutoGPT↳ Platform for buil ...

LangChain. MCP. OpenAI GPT Actions. CrewAI. Eliza.Orbs Agentic plugs into the agent frameworks developers already use. If your agent needs on chain execution, it picks up the Orbs tool like any other capability.No custom integrations. Just tools. ...

Core Insights - The era of simply wrapping AI with APIs and prompts is over, as AI applications transition from "one-time generation" to "continuous execution" [2] - The software infrastructure is being rewritten, with frameworks becoming more important than models, as highlighted by LangChain's recent developments [3][4] - The future of AI will focus on the core components of modern agents: system prompts, planning tools, sub-agents, and file systems [18][27] Group 1: Evolution of AI Agents - The capabilities of AI agents have significantly improved, moving from simple models to more complex systems that can run in loops and call tools effectively [7][10] - The development trajectory of agents shows that initial concepts have evolved into frameworks that enhance predictability and reliability [8][10] - The distinction between single agents and collaborative multi-agent systems will be crucial, with communication being a key factor in their effectiveness [9][11] Group 2: Framework vs. Model - The debate on whether models will dominate frameworks or vice versa suggests that frameworks will ultimately be more critical, as they enable models to be utilized effectively [14][15] - Frameworks serve as the interaction layer between models and their environments, providing essential tools for agent development [16][17] Group 3: Core Components of Modern Agents - The four core components of modern agent architecture are system prompts, planning tools, sub-agents, and file systems, which facilitate better management of context and tasks [27] - System prompts act as standard operating procedures for agents, guiding their actions from the moment they are activated [20] - Planning tools help agents generate and manage task lists, while sub-agents allow for context isolation and task delegation [21][22] Group 4: Memory and Context Management - Memory types in agents include semantic memory, episodic memory, and procedural memory, which define how agents learn and adapt over time [38] - Context compression techniques are essential for managing large amounts of information, ensuring that agents can operate efficiently without overwhelming their processing capabilities [32][34] Group 5: Future Directions and Commercialization - LangChain's future focus will be on enhancing observability and building a comprehensive platform for agent development, following a recent $125 million funding round [61][63] - The emphasis on tools, instructions, and skills will remain the primary differentiators for companies in the AI space, as frameworks and models become more standardized [64]

Core Viewpoint - OpenClaw represents a significant advancement in AI agent frameworks, enabling a transition from "dialogue AI" to "execution AI," allowing users to issue commands through instant messaging software for autonomous task completion, which could reshape the financial industry's AI applications and address current challenges in AI utilization [3][10]. Group 1: Development and Features of OpenClaw - OpenClaw is an open-source AI agent framework that allows for real-world interaction by connecting with existing applications, enabling tasks such as coding, email management, and file organization without human intervention [8][10]. - The framework addresses previous limitations of AI agents, such as lack of cross-scenario applicability and memory management, by providing a standardized execution environment and a "heartbeat" mechanism for proactive task management [9][10]. - OpenClaw's ability to autonomously operate various software applications marks a departure from earlier AI frameworks, which required human input for execution [10]. Group 2: Impact on Financial Industry - The financial sector has shown a keen interest in AI, initially focusing on proprietary large models but shifting towards AI agent development as the capabilities of large models have improved [13][14]. - Financial institutions are increasingly embedding large models into existing workflows, automating repetitive tasks and developing AI agents tailored to specific operational needs [14][15]. - Major banks like Morgan Stanley and JPMorgan Chase have launched platforms that utilize AI agents for various functions, including investment analysis and contract generation, demonstrating the growing integration of AI in financial services [15][16]. Group 3: Challenges and Future Directions - Despite the potential of OpenClaw, challenges remain, including the inherent limitations of large models and the uncertainty in AI agent execution, which can lead to errors in task completion [17][18]. - The transition from passive to proactive decision-making in financial AI agents is essential, with OpenClaw's framework supporting goal-driven decision-making rather than rule-based execution [18][19]. - Financial institutions are encouraged to redesign business processes to be AI-native, enhancing the overall value derived from AI applications and improving data governance to leverage proprietary data effectively [21][22].

In this video, we'll be showing how to use link chain and linksmith skills with cloud code. We'll see how Cloud Code can use skills to effectively build, observe, and iterate on an agent. To start, we just ask Cloud Code to generate us a deep research agent that can delegate tasks to workers.You can see Cloud call our skills for guidance. Without them, Cloud wouldn't know best practices. With skills, it knows dependencies, structures, and orchestration out of the gate.And we can watch it put together our ag ...

Core Insights - The article emphasizes the transition from traditional chatbots to Long-horizon Agents, which are capable of performing complex tasks over extended periods, thus redefining the value proposition of AI products from speed of response to quality of output [3][8][10]. Group 1: Long-horizon Agents - Long-horizon Agents are designed to operate autonomously over longer time spans, allowing for multi-step decision-making and iterative processes, which are essential for tasks like research reports and code reviews [16][20]. - The emergence of Long-horizon Agents marks a significant shift in AI capabilities, moving from simple question-answer interactions to producing high-quality deliverables that require time and context [7][8][11]. Group 2: Harness Concept - The concept of "Harness" is introduced as a runtime environment that includes best practices for building Long-horizon Agents, distinguishing it from traditional frameworks by providing integrated tools and capabilities [11][23]. - Harnesses facilitate the development of agents that can autonomously manage tasks, including planning, memory management, and sub-task coordination, thus enhancing their effectiveness [11][23][24]. Group 3: Evolution of AI Agents - The evolution of AI Agents is categorized into three phases: simple prompting and chaining, cognitive architecture, and the current Long-horizon Agent era, which began around mid-2025 [26][30][31]. - The transition to Long-horizon Agents is characterized by improved model capabilities and a focus on context engineering, which is crucial for optimizing agent performance [29][34]. Group 4: Applications and Future Directions - Long-horizon Agents are particularly effective in generating initial drafts for various applications, such as coding, research, and customer support, where they can significantly reduce the workload for human users [20][22]. - The future of AI development is expected to focus on enhancing context engineering, memory management, and the integration of file systems, which are seen as critical components for the success of Long-horizon Agents [34][42][46].

Core Insights - OpenClaw, a new desktop agent, has gained significant popularity in tech communities, enabling users to interact with AI in a natural chat interface and perform complex tasks [1][4] - The product's design allows for local operation, enhancing privacy and security, which has led to a surge in demand for MacMini as a dedicated device for running OpenClaw [4][17] - The rise of OpenClaw has sparked interest from cloud service providers, with companies like Alibaba Cloud and Tencent Cloud quickly launching dedicated services and templates for OpenClaw deployment [4][5] Group 1: Product Features and Market Response - OpenClaw can perform a variety of tasks, from comparing car dealership quotes to managing email subscriptions and flight bookings, showcasing its versatility [1] - The agent's ability to maintain long-term memory and context allows it to proactively send reminders and alerts, likened to a "24-hour standby Jarvis" [1] - The rapid adoption of OpenClaw has led to a proliferation of tutorials and guides, indicating strong community engagement and interest [1][4] Group 2: Financial Implications and Operational Costs - Users have reported high operational costs associated with OpenClaw, particularly due to its extensive use of API tokens, which can be quickly depleted [5][6] - Traditional chatbots consume fewer tokens per interaction compared to OpenClaw's autonomous operation, which can lead to significant expenses [6] - The need for efficient and cost-effective models is emphasized, as the performance of agents like OpenClaw heavily relies on underlying large models [6][7] Group 3: Competitive Landscape and Future Trends - The emergence of OpenClaw has intensified competition among desktop agents, with various players entering the market, including Manus and MiniMax [8] - The future of software is shifting towards a "thousand-end battle," where the focus will be on the capabilities of agents rather than just models [8] - Major tech companies like Apple and Microsoft are expected to evolve their AI offerings into comprehensive agents, leveraging their unique system-level access [10][11][12] Group 4: Hardware and Infrastructure Developments - The demand for dedicated hardware, such as MacMini, has surged due to its compatibility with OpenClaw, although it is not seen as a long-term solution [17][18] - New hardware solutions, including AI mini PCs and cloud-based "AI boxes," are emerging to provide cost-effective alternatives for users needing lightweight agents [20] - The competition for control over desktop agents is expected to intensify, with both software and hardware players vying for market share [20]

Core Viewpoint - The emergence of "long-horizon agents" is reshaping the software engineering paradigm, moving from deterministic code-based systems to models that operate as black boxes, requiring real-time execution to understand their behavior [2][3][6]. Group 1: Long-Horizon Agents - Long-horizon agents are seen as a turning point in AI, with predictions that their adoption will accelerate by the end of 2025 to 2026 [2]. - These agents function more like "digital employees," capable of executing tasks over extended periods, learning from trial and error, and self-correcting [2][3]. - The transition to long-horizon agents may challenge traditional software companies, similar to the shift from on-premises to cloud solutions, where not all companies successfully adapted [2][3]. Group 2: Differences in Software Development - Traditional software development relies on deterministic logic written in code, while agent-based systems introduce non-deterministic behavior, making it necessary to observe their real-time execution to understand their operations [30][32]. - The concept of "tracing" has become crucial in agent systems, allowing developers to track internal processes and understand the context at each step, which differs significantly from traditional software debugging methods [31][32]. - The iterative process of developing agents is more complex, as developers cannot predict behavior before deployment, necessitating more rounds of refinement and adjustments [34][36]. Group 3: The Role of Data and Instructions - Existing software companies possess valuable data and APIs that can be leveraged in the agent era, but the ability to effectively utilize these assets will depend on new engineering approaches [37][38]. - The instructions on how to use data effectively are becoming increasingly important, as traditional methods of human execution are being automated through agents [38]. - The integration of domain-specific knowledge into agent systems is essential for their effectiveness, as seen in examples from the financial sector [38]. Group 4: Future of Agent Development - Memory capabilities in agents are anticipated to become a significant competitive advantage, allowing them to learn and improve over time [51][52]. - The development of user interfaces for long-horizon agents will likely require both synchronous and asynchronous management to handle tasks effectively [53][54]. - Code sandboxes are expected to become a critical component of agent capabilities, enabling safe execution and verification of scripts [56].