OpenAI Updates Agents SDK With Sandboxing And Frontier Model Harness@OpenAI rolled out an Agents SDK update aimed at enterprises building their own AI agents on top of its models. The headline additions are sandboxing, which lets agents run in siloed workspaces with limited file and code access, and a new in-distribution harness for frontier models that pairs the agent with approved tools.The pitch is safer long-horizon agents. Running agents unsupervised is a known risk because they go off-script in unpred ...

OpenAI updates its Agents SDK to help enterprises build safer, more capable agents https://t.co/o4WU6sAVvN ...

Core Insights - The article discusses the transformative impact of AI, particularly OpenAI's Codex, on software engineering, highlighting a shift from traditional coding to a more management-oriented role for engineers [3][5][12]. Group 1: AI Integration in Software Development - Currently, 95% of OpenAI engineers use Codex, with 100% of code merge requests being reviewed by it, indicating a significant reliance on AI for coding tasks [5][7]. - Engineers are evolving into technical leaders, managing multiple AI agents rather than writing code directly, which reflects a paradigm shift in software development [5][13]. - The expectation is that within the next 12 to 18 months, AI models will be capable of executing complex tasks over several hours, fundamentally changing the nature of software products [5][12]. Group 2: Efficiency and Productivity Gains - Engineers who utilize Codex submit 70% more pull requests (PRs) compared to those who do not, demonstrating increased productivity through AI tools [7][20]. - Codex has automated code reviews, reducing the time required for this task from 10-15 minutes to just 2-3 minutes, allowing engineers to focus on more engaging work [20][22]. Group 3: Evolution of Engineering Roles - The role of engineers is shifting towards that of managers who oversee AI agents, requiring new skills to ensure effective collaboration with AI [12][13]. - The metaphor of engineers as "wizards" using "spells" (code) to command AI reflects the growing complexity and capability of AI tools in software development [14][15]. Group 4: Challenges and Best Practices - A team at OpenAI is experimenting with a codebase entirely written by Codex, facing challenges in ensuring the AI meets specific requirements without a fallback option [18][19]. - Successful AI deployment in organizations requires both top-down strategic support and bottom-up employee engagement to foster a culture of AI utilization [44][45]. Group 5: Future of Startups and Entrepreneurship - The concept of "one-person billion-dollar startups" suggests that individuals leveraging AI tools can achieve significant productivity, potentially leading to a surge in small startups [30][31]. - The article posits that as software development becomes easier, there may be a proliferation of small companies, leading to a B2B SaaS boom [31][32]. Group 6: Management Philosophy and AI - The management philosophy emphasizes empowering high-performing employees, akin to supporting a lead surgeon in an operating room, to enhance productivity and innovation [39][40]. - AI tools can assist managers in predicting potential bottlenecks and proactively addressing issues, thereby improving team efficiency [41][42]. Group 7: AI Deployment Challenges - Many companies face negative ROI from AI projects due to a disconnect between management directives and employee capabilities, highlighting the need for a dedicated "AI task force" to bridge this gap [43][44]. - The ideal "AI task force" should consist of technically inclined individuals who are not necessarily software engineers but can effectively leverage AI tools [46].

Core Insights - OpenAI's internal data reveals that 95% of its engineers are using Codex for programming, with 100% of pull requests (PRs) being reviewed by Codex, indicating a significant shift in software engineering practices [4][9][19] - The company is experimenting with a team maintaining a codebase entirely written by Codex, which could fundamentally change development methodologies [4][12] - Engineers are evolving from traditional coding roles to managing multiple AI agents, likening their work to that of "wizards" casting spells to accomplish tasks [5][6][10] Group 1: AI Integration and Impact - The deep integration of AI tools has led to engineers who use Codex generating 70% more PRs than those who do not, with this gap widening over time [4][18] - OpenAI emphasizes the need for developers to build for the future capabilities of AI models rather than their current state, as many existing scaffolding solutions may become obsolete [4][14][15] - The company views itself as an ecosystem platform aimed at enhancing the overall landscape rather than stifling startups through competition [8] Group 2: Future of Software Engineering - The next 12 to 24 months are expected to see AI models capable of executing complex tasks for several hours, marking a significant advancement in AI capabilities [7] - The rise of "one-person billion-dollar startups" is anticipated, with a corresponding increase in small SaaS companies catering to these individuals, potentially transforming the venture capital ecosystem [7][43] - The emergence of a B2B SaaS golden age is predicted, where the ease of software creation will lead to a proliferation of micro-companies [7][43][44] Group 3: Management and Workforce Dynamics - As AI tools enhance productivity, top performers are expected to leverage these tools to achieve greater efficiency, leading to a wider distribution of team productivity [36][37] - Management roles are evolving, with leaders spending more time supporting top performers and ensuring they have the resources needed to excel [37][41] - The integration of AI tools is likely to enable managers to oversee larger teams, similar to how engineers manage multiple AI agents [38][39]

但 LangChain 创始人 Harrison Chase 对于 OpenAI 在文中的一些观点持有异议，尤其是「通过 LLMs 来主导 Agent」的路线，迅速发表了一篇长文回 应。 Harrison Chase 认为，并非要通过严格的「二元论」来区分 Agent，目前我们看到大多数的「Agentic 系统」都是 Workflows 和 Agents 的结合。理想 的 Agent 框架应该允许从「结构化工作流」逐步过渡到「由模型驱动」，并在两者之间灵活切换。 相比 OpenAI 的文章，Harrison Chase 更认同 Anthropic 此前发布的如何构建高效 Agents 的文章，对于 Agent 的定义，Anthropic 提出了「Agentic 系 统」的概念，并且把 Workflows 和 Agents 都看作是其不同表现形式。 总的来说， 这是大模型派（Big Model）和工作流派（Big Workflow）的又一次争锋， 前者认为每次模型升级都可能让精心设计的工作流瞬间过 时，这种「苦涩的教训」让他们更倾向于构建通用型、结构最少的智能体系统。而以 LangGraph 为代表的后者，强调 ...

微信搜一搜也推出 "附近的工作" 专区，用户在微信搜索 "附近的工作""兼职""日结""零工" 等关键 词，即可看到零工岗位的合集。这些信息由腾讯对本地招聘中介机构发布的零工招聘信息智能梳理 而来。目前该项服务已覆盖全国超 200 个地市。 去年 10 月，微信即针对零工市场推出找工作服务，上线初期仅覆盖广东江门。定位江门的用户， 可在微信进入 "零工市场"。"零工市场" 接入微信支付的 "微工卡" 数字化解决方案，提供微信零钱 及银行卡结算方式，并联合腾讯电子签、腾讯微保等业务，提供合同签约、实时定位打卡、日结保 险等服务。 蜜雪冰城利润增加、利润率提高。 微信要帮灵活就业者找工作。 微信上线 "附近的工作" 小程序，用户可在小程序内，看到附近的零工工作岗位及薪酬。目前该服 务已覆盖广东江门、河北唐山、江西九江、广东深圳、北京朝阳等地。腾讯称已提供超过 2 万个岗 位，所有岗位信息均由政府人社部门审核发布。 3 月 26 日，蜜雪冰城发布上市后首份财报。集团去年营业收入 248.3 亿元，同比增长 22.3%，净利 润 44.5 亿元，同比增长 39.8%。全球门店总数截至去年底 46479 家。上市 20 ...

Investment Rating - The report maintains a rating of "Outperform the Market" for the artificial intelligence sector [3]. Core Insights - Baidu officially launched the Wenxin large model 4.5 and Wenxin model X1 on March 16, 2025, which are now available for free on the Wenxin Yiyan website. The Wenxin model 4.5 is a native multimodal large model with superior multimodal understanding and generation capabilities, outperforming GPT-4.5 in tests, and its API call price is only 1% of GPT-4.5's [5][10]. - Google's Gemini 2.0 integrates native image generation technology, allowing users to generate images through natural language commands, creating a competitive edge against OpenAI. Its features include text-image integration, conversational image editing, and enhanced reasoning capabilities [23][41]. - The upcoming NVIDIA GTC 2025 conference is highlighted as an event to watch for new information and investment opportunities, including next-generation AI chips and cooling technologies [5]. Industry Changes - Baidu's Wenxin model 4.5 and X1 were launched on March 15, 2025, and are now available on the Wenxin Yiyan website for free. The models will be integrated into various Baidu products, enhancing user experience [8][41]. - The report emphasizes the advancements in AI models, particularly the capabilities of Wenxin model 4.5 in multimodal understanding and the innovative features of Gemini 2.0 [10][23]. Continuous Focus Targets - Key companies to watch include NVIDIA, AMD, and Haiguang Information in the GPU sector; Anlu Technology in FPGA; Qualcomm and Allwinner Technology in SoC; iFlytek in natural language processing; and companies like Desay SV and Zhongke Chuangda in autonomous driving [26]. Market Review - The AI index (CSI) decreased by 1.98% in the last period but has increased by 9.39% since the beginning of the year. The Shanghai Composite Index rose by 1.39%, and the CSI 300 Index increased by 1.59% [27].

Investment Rating - The report assigns an "Overweight" rating for the industry, consistent with the previous rating [2]. Core Insights - OpenAI has launched four AI agent toolkits, including a new Responses API and open-source Agents SDK, which simplify the development of AI agent applications and enhance product development capabilities [4]. - The introduction of new tools accelerates the diverse application of agents, responding to competitors like Google, Microsoft, and Alibaba, while reinforcing OpenAI's competitive edge established by its first AI agent product, Operator [9]. - Recommended stocks include Dingjie Zhizhi, Foxit Software, and iFlytek, with beneficiaries being Fanwei Network, Maifushi, and Rundamedical [9]. Summary by Sections Investment Highlights - The Responses API features web and document search tools, with the web search tool achieving higher accuracy than competitors, significantly reducing manual search costs [9]. - The computer operation tool, supported by the CUA model, automates tasks like clicking and scrolling, enhancing agent interaction with the external world [9]. - The open-source Agents SDK improves agent collaboration, allowing for customizable LLM models and tools, which simplifies multi-threaded task processing and reduces developer workload [9]. - OpenAI employs a "free API + paid tools" business model, expanding its market through free offerings while recouping R&D costs via paid tools, ensuring a sustainable financial foundation for future growth [9].