OpenClaw，是当下最火的开源个人 AI 助手。很多人不知道的是，OpenClaw 背后，核心是一个极简框架 Pi-coding-agent。 在 OpenClaw 的系统架构中，Pi agent 是 Gateway 控制层的核心子系统，控制了所有 agent 的推理和工具调用。 ⬆️ 关注 Founder Park，最及时最干货的创业分享 超 22000 人的「AI 产品市集」社群！不错过每一款有价值的 AI 应用。 邀请从业者、开发人员和创业者，飞书扫码加群： 进群后，你有机会得到： 01 和 Claude Code、Cursor、Codex 不同的是，Pi 最大的特点是「做减法」：系统提示词和工具定义加起来不到 1000 tokens，核心只有 read、write、 edit、bash 四个工具，没有内置 plan mode，没有 to-do 系统，没有 MCP 支持，没有权限弹窗，甚至没有绑定任何特定模型。 但就是这样一个「什么都没有」的框架，在 Terminal Bench 2.0 上与 Codex、Cursor、Windsurf 一同排进了前五。在 GitHub 上，Pi 积累了超过 240 ...

Core Insights - The article discusses the evolution of AI development practices from Prompt Engineering to Context Engineering, and now to Harness Engineering, emphasizing the importance of the environment in which AI agents operate [4][40][41] Group 1: Evolution of Engineering Practices - In 2023, Prompt Engineering was at its peak, focusing on crafting effective prompts for AI to deliver results [9] - By mid-2025, Context Engineering emerged, shifting the focus to designing dynamic systems that provide the necessary context for AI tasks [9][10] - As of February 2026, Harness Engineering was introduced, highlighting that the environment in which AI agents operate is crucial for their performance [11][12][13] Group 2: OpenAI's Experiment and Findings - OpenAI conducted an experiment with a team of engineers who delivered over 1 million lines of code without writing any human code, relying entirely on Codex Agent [15] - The experiment revealed that the most significant challenges lie in designing the environment, feedback loops, and control systems for AI agents [22][42] - The team learned that a well-structured documentation system is essential, evolving from a single large document to a more organized directory structure [17][18] Group 3: Framework of Harness Engineering - Birgitta Böckeler outlined a three-dimensional framework for Harness Engineering, which includes Context Engineering, Architectural Constraints, and Entropy Management [24][25][26] - Context Engineering ensures that agents receive the right information at the right time, while Architectural Constraints enforce boundaries through automated mechanisms [24][25] - Entropy Management addresses the degradation of the system over time, ensuring that the harness remains effective and does not become outdated [26] Group 4: Industry Adoption and Examples - Companies like Stripe and LangChain are implementing Harness Engineering principles, with Stripe's Minions system merging over 1,300 AI-generated pull requests weekly [28][29] - LangChain demonstrated a significant performance improvement in its coding agent by optimizing the harness without changing the underlying model [29][30] - The concept of Harness Engineering is being internalized by tool vendors, with MCP (Model Control Protocol) becoming a standard for agent tool access [31] Group 5: Future Directions for Engineers - The core responsibilities of engineers are shifting from writing code to designing environments that ensure reliable operation of AI agents [33] - Engineers are now focused on building documentation systems, defining business intents in machine-readable formats, and creating automated validation mechanisms [33][34] - The industry is recognizing the need for a deeper understanding of system design over mere coding speed, leading to a re-evaluation of team structures and roles [35][36]

Hey everyone, welcome back to OpenAI Build Hours. I'm Christine. I'm on the startup marketing team and today I'm joined with Charlie and Ryan.>> Hey folks, how's it going. >> Hey everybody. >> Awesome.Um, so Charlie is on our Dev X team and he will be leading the session and Ryan actually came all the way from Seattle to be with us live in the studio today. Um, and he's going to be chatting about the future of work. Um so today's session is all about um API and codecs.And if this is your first build hour, t ...

Core Insights - OpenAI's engineers have significantly reduced their coding tasks, relying instead on Codex to generate code autonomously [2][3] - This shift reflects a broader cultural change within OpenAI, emphasizing a bottom-up approach to innovation and project development [5][7] Group 1: Engineering Culture and Process - OpenAI maintains a strong startup culture, allowing engineers high autonomy and quick decision-making [5] - The development of Codex was driven by a small team that rapidly transitioned from concept to deployment in just seven weeks [8] - The bottleneck in the coding process has shifted from code generation to quality assurance, prompting engineers to rethink their roles [9][10] Group 2: Harness Engineering Concept - The new approach, termed "Harness Engineering," involves engineers acting as "capability architects" rather than traditional coders [13][14] - Engineers focus on designing environments, feedback loops, and architectural constraints, allowing agents to execute tasks [11][12] - The project began with a blank code repository, where Codex autonomously generated the initial architecture and configurations [15] Group 3: Enhancing Agent Capabilities - Engineers are tasked with making applications AI-readable, ensuring that agents can interact effectively with user interfaces [19][20] - Implicit knowledge must be codified into the codebase, allowing agents to access necessary information during execution [21][22] - A structured architecture is essential for AI efficiency, with strict hierarchies and dependencies enforced [26][27] Group 4: Quality and Maintenance - Engineers translate aesthetic preferences into coding standards, ensuring that AI-generated code adheres to human taste [29][30] - A "garbage collection" mechanism is implemented to clean up suboptimal code generated by AI, preventing technical debt accumulation [32][34] - The process of code generation and testing has become automated, with agents now capable of self-testing and debugging [45][54] Group 5: Future Implications - The shift towards agent-driven development processes indicates a potential transformation in software engineering, focusing on system design and feedback mechanisms [59] - OpenAI's model of "agents writing code while humans design systems" is still in exploratory phases, with ongoing challenges regarding long-term code management and human oversight [59]