智东西（公众号：zhidxcom）
作者 | 毕伟豪
编辑｜李水青
智东西10月31日报道，今天，快手在技术沙龙上结合企业级AI生产力平台iWork的实践，分享了AI如何从辅助编程逐步走向团队协作与组织级提效，并沉淀为企业生产力基础设施。
快手企业智能体建设负责人、企业AI生产力专家沈浪复盘了快手过去两年AI生产力建设的四个阶段：从AI赋能工程师、全体员工，到AI赋能团队和组织。他特别提出，个人提效并不等于团队提效，企业需要将个人经验沉淀为可复用的数字员工、工作流和Skill，进一步推动团队能力共享与组织级Agent协作。
《软件工程3.0》作者、同济大学特聘教授朱少民则从工程方法论出发，分享了企业级智能体落地的关键路径，强调通过规约工程、驾驭工程和质量门禁，建立Agent可靠交付的工程体系。
快手磁力引擎风控技术负责人王东旭则从组织变革的角度，分享了快手从“超级个体”走向“液态组织”和“臭鼬工厂”的探索，尝试通过打破岗位边界、减少协作环节，让AI带来的技术效率真正转化为组织效率。
OPPO产研提效产品负责人郝刚分享了OPPO内部的AI Native演进路径，介绍了从工具普及、个人提效到组织提效的实践，以及如何通过高质量知识库、AI交付工作流和智能体协同交付平台CoMind，推动AI融入完整研发流程。
快手磁力引擎风控质量负责人吕冬冬则聚焦AI规模化应用中的质量保障问题，他认为，AI提升研发单兵产出速度的同时，也让验证能力成为新的瓶颈。
面对需求变更频繁、AI生成用例可信度不足等挑战，团队正在探索需求测试智能分级、风险挖掘和智能体自主测试，并通过异源交叉、人工真值校准等方式，解决“AI测AI”过程中的可信度问题。
一、个人、团队再到组织，AI重塑企业组织模式
当AI工具逐渐普及，企业面临问题，是如何将分散在员工个体身上的效率提升，转化为团队协作和组织能力。来自快手、OPPO的三位嘉宾分享了企业AI转型从个人提效走向组织变革的不同路径。
1、沈浪：从个人到组织，快手AI生产力的四阶段演进
快手企业智能体建设负责人、企业AI生产力专家沈浪认为，企业AI转型需要经历从个人到团队、再到组织的逐步演进。快手将这一过程划分为四个阶段：AI赋能工程师、AI赋能全体员工、AI赋能团队，以及AI赋能组织。
第一阶段，快手主要通过产品升级优化研发范式，让AI从代码辅助逐步走向Agent自主研发。第二阶段，快手将AI使用范围从工程师扩展到全体员工，让不同岗位的员工都能充分使用AI。
第三阶段是快手实践中的一个重要转折点。 通过大量实践，快手将技术团队划分为四类：复杂业务开发团队、AI友好开发团队、产研闭环交付团队和AI原生交付团队。
四类团队在AI研发流程中的参与程度和自主化程度各不相同，AI也由辅助单项研发任务，逐步向承担完整研发交付流程演进。
在非技术团队中，快手则提出了“先蒸馏、再折叠”的提效思路：先将个人经验AI化，通过数字员工、可复用工作流等方式沉淀个人能力，再通过知识、Skill和工具共享，将个人经验转化为团队可以共同调用的能力。
基于这一思路，快手推出MyFlicker 2.0团队AI产品形态。据沈浪介绍，目前快手园区已有1000多个团队Agent在运行。
第四阶段，快手开始探索从团队AI走向组织AI。 沈浪指出，团队AI的规模和成熟度各不相同，而整个组织是一个全新的业态，虽然其运行逻辑与团队AI相似，但在协作方式和能力整合上存在本质区别。
以HR体系为例，组织级AI不仅需要处理招聘、绩效、员工服务等单项任务，还需要打通不同职能之间的知识、数据和业务流程。
为此，快手正在研发MyFlicker 3.0，通过增加工作台功能，为组织AI提供稳定的Agent底座，推动AI从团队内部的能力复用进一步走向跨团队、跨业务的组织协作。
2、王东旭：超级个体、液态组织再到臭鼬工厂，快手探索组织变革新模式
快手磁力引擎风控技术负责人王东旭认为，AI对人的影响呈现出一条“微笑曲线”：能力强的人借助AI进一步放大优势，能力相对弱的人则可能被拉开差距，马太效应愈发明显。
面对这一变化，快手探索了三个阶段：从打造“超级个体”，到构建“液态组织”，再到形成“臭鼬工厂”模式。
第一阶段，快手通过IDE、浏览器Vibe Coding和CLI Agent等方式，不断放大个体能力。但超级个体并不能自动带来团队效率提升。
为此，快手推动算法与工程、前端与后端融合，让具备业务上下文的技术人员承担更多工作，也通过技术让渡，将部分提示词工程交给运营人员。
在前后端融合过程中，快手还建设了多仓全栈Agent Harness。不过，即使实现AI全栈开发，组织效率仍有天花板。
王东旭指出，开发环节约占端到端交付流程的20%，即使AI让代码开发效率提升10倍，整体交付效率也只能提升约22%。
因此，快手最终选择了“消除”这条组织提效路线，通过绑定共同业务目标、减少流转节点、集中产运研决策等方式，消除不必要的协调成本，让小团队形成业务闭环，逐步形成“臭鼬工厂”模式。
据王东旭介绍，经过这套组织与工程方式的调整，团队需求吞吐率提高了133%，交付周期缩短了79%，全栈交付需求数提高了800%，并获得快手内部领域级AI样板间的荣誉。
不过，臭鼬工厂模式也面临新的挑战：当小团队需要与大型组织协作时，外部流程和跨部门协调仍可能形成阻碍。
王东旭认为，这一瓶颈的打破需要更多臭鼬工厂的出现，通过多个具备自主交付能力的小团队协同，逐步打破大型组织中的协作壁垒。
3、郝刚：走向AI Native组织，需要自上而下和自下而上
OPPO产研提效产品负责人郝刚分享了OPPO内部的AI Native演进路径：从工具普及、个人提效，到组织提效，最终走向AI Native组织。
在工具普及阶段，OPPO将AI引入编程、测试、运营、审核、游戏和设计等日常工作，甚至曾要求员工达到一定的月度Token消耗量。同时，OPPO通过AI README，让项目文档和工作内容更容易被AI理解，并逐步搭建数字人研发、模型广场与知识库，以及Skill和MCP中心等平台能力。
随着AI工具逐渐普及，OPPO也发现，个人效率提升并不意味着团队交付效率同步提高。为此，OPPO开始提高研发流程中AI能够参与和掌控的环节占比，并重点推进两件事：建设高质量知识库，以及打造能够持续演进的AI交付工作流，以扩大AI协同交付的覆盖范围，减少重复交接和等待。
在组织层面，OPPO先选择部分人员开展AI转型试验，再逐步推进业务流程重构、专业能力沉淀和平台底座建设。其内部的CoMind智能体协同交付平台，能够承接从想法、需求提出到最终上线交付的完整流程，为团队协作提供AI能力支持。
对于如何走向AI Native，郝刚认为，需要自上而下与自下而上两条路径共同推进：管理者既要亲自上手，也要判断转型方向；员工则需要从超级个体走向协同的“蜂群”，让判断、执行和经验彼此连接，形成团队复利。
二、AI无所不在，企业级AI需要保障可信
当AI深入团队协作和企业研发流程，如何让Agent在复杂业务环境中稳定执行任务，如何衡量交付质量，又如何防止效率提升带来新的风险，这些问题成为了新的挑战。两位嘉宾分别从方法论和实践两个层面分享了，企业该如何解决这些问题。
1、朱少民：企业级智能体落地，需要精密的工程体系
《软件工程3.0》作者、同济大学特聘教授朱少民认为，企业级智能体落地不能只关注模型能力和代码生成速度，更需要建立一套可靠的工程体系。
企业生产力应当以可交付的客户价值衡量，即企业生产力=可交付客户价值/总投入。如果效率提升伴随着质量、可交付率和可治理性的下降，企业整体生产力反而可能降低。
他指出，生成不等于完成、个人提效不等于组织提效、Demo不等于可靠生产，企业需要将关注点从生成速度转向交付速度，从个人效率转向组织吞吐量，从模型能力转向工程系统。
在具体落地方法上，朱少民重点介绍了规约工程与驾驭工程。规约工程旨在定义Agent的工作目标、上下文、约束和验收条件。驾驭工程（Harness Engineering）则通过系统化的工程控制，为Agent建立明确的执行边界和运行环境。
此外，企业还需要建立质量门禁。由于Agent的行为具有概率性，不能仅依赖其自身判断任务是否完成，而应通过分层质量检查，对产出进行验证、拦截和反馈，形成持续改进的闭环。
对于企业的演进路径，朱少民提出，可以分三步推进：首先进行架构清理和知识沉淀，为Agent参与研发打好基础；其次搭建Harness框架，建立Agent自主执行任务的工程环境；最后构建质量门禁和反馈闭环，让Agent从依赖外部能力和人工指导，逐步走向系统自主。
2、吕冬冬：构建合理需求测试分级体系，让AI更好地测试AI
快手磁力引擎风控质量负责人吕冬冬认为，AI正在显著提升研发单兵的交付速度，但代码生成能力的提升，并不意味着质量保障能力同步增长。
当需求变更更加频繁、代码产出更加细碎，甚至出现一些表面正确、实际语义错误的隐性缺陷时，验证能力就可能成为新的瓶颈。
这一变化也带来了新的质量风险：测试速度跟不上研发节奏，项目容易堆积在质量验证环节；AI生成的测试用例和结论缺乏足够的可信度；重复代码增多、代码腐化加速，也让质量审核更加困难。
此外，团队对AI提效的主观感受可能强于实际产出，容易造成过于乐观的判断。
针对这些问题，吕冬冬分享了快手团队正在探索的需求测试智能分级体系。其团队将人工判断经验与AI分析能力结合，建立智能分级流程，把需求划分为高、中、低三个风险等级，并匹配不同的测试策略：高风险需求由人工主导、AI辅助；中风险需求由AI与人工协同；低风险需求则更多交由AI主导。
分级过程中还引入了包含12类检测项的风险挖掘体系，覆盖核心业务链路改动、数据安全与合规、上下文依赖及复杂度等因素，同时保留人工最高决策权限。
随着AI逐渐承担更多测试工作，质量团队也需要重新划分职责边界。吕冬冬提出，智能测试要让Agent逐步参与决定“测什么、怎么测、测多深”，通过策略决策模型，根据风险匹配测试组合，并给出可解释的策略依据，最终形成结构化测试计划。
在测试执行环节，则需要进一步实现多工具协同、自主执行、稳定可靠和降本提效，让测试从人工操作逐步走向智能体驱动。
不过，让AI参与测试，并不意味着可以直接相信AI给出的结论。尤其在“AI测AI”的场景中，被测对象本身具有不确定性，测试工具的判断也可能存在偏差。
为此，快手团队探索了几种可信度保障方法：通过异源交叉验证降低单一模型偏差，以变异测试检验测试能力，用人工真值校准AI判断，并坚持依据证据而非结论做质量决策。
结语：从个人提效到组织协同，AI深度影响企业协作模式
从个人使用AI编写代码，到团队借助智能体协同交付，再到企业重新设计组织和研发流程，AI带来的变化正在从单点工具提效，深入到企业生产力体系的重构。
企业引入AI，最终要看的不只是代码写得有多快、单个员工效率提升多少，更要看能否让团队协作更顺畅、交付更加可靠，并将个人经验沉淀为组织可以持续复用的能力。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Intellectual (WeChat Official Account:zhidxcom)
The author.
Edit Qi Shui Qing
On October 31st, Ji-Man reported that today, fast-trackers, combining the practices of the Enterprise-level AI Productivity Platform iWork, shared how AI moved from supporting programming to team and organization-level effectiveness and settled into an enterprise productivity infrastructure.
AI agent Construction Manager, Enterprise AI Productivity Specialist, reset four stages of AI productivity construction in the past two years: from AI Enabling Engineers, to all employees, to AI Enabling Teams and Organizations. In particular, he argued that personal effectiveness did not amount to team effectiveness and that enterprises needed to sink their personal experience into reusable digital staff, workflow and Skill, further promoting team capacity-sharing and organizational-level Agent collaboration.
The author of Software Project 3.0 and special professor Zhu Xiaomin, at the University of Toji University, shared the key path to the landing of AI agent at the enterprise level, from the methodological point of view of engineering, emphasizing the establishment of an engineering system for the reliable delivery of Agent by regulating engineering, managing engineering and quality doors.
Wang Dong Wuk, Director of Wind Control for Rapid Magnetic Engines, shared his exploration of moving from a “super-individual” to a “liquid organization” and a “skunk factory” from the perspective of organizational change, trying to transform the technological efficiency of AI into organizational efficiency by breaking the job boundaries and reducing the coordination links.
Ho Hao, the producer of the OPPO yield product, has just shared the AI Native evolution path within the OPPO, describing practices that range from universal access to tools, individual effectiveness to organizational effectiveness, and how to facilitate the integration of AI into the full R & D process through a high-quality knowledge base, AI delivery of Z52Z and Z42Z collaborative delivery platform CoMind.
The head of wind quality control for fast-hand magnetic engines, Lu winter, focused on quality assurance in AI scale applications, which, in his view, while increasing the speed of R & D individual output, also created new bottlenecks in certification capacity.
In the face of challenges such as frequent changes in demand and insufficient credibility in the use of AI-generated examples, the team is exploring the need-testing smart rankings, risk mining and AI agent autonomous testing, and addressing the credibility of the “AI-testing” process through the intersection of sources and the calibration of artificial values.
I. Individuals, teams back to the organization, AI reshaping the business organization model
As AI tools become available, enterprises face the challenge of transforming the efficiency gains that are dispersed among individual employees into teamwork and organizational competencies. Three guests from the Express, OPPO, shared the different paths of corporate AI transition from individual effectiveness to organizational change.
1 Shen waves: from individual to organizational, fast-track AI productivity four stages
AI agent construction manager and corporate AI productivity expert Shizun believes that the transformation of the firm requires a gradual evolution from individual to team to organization. The fast-tracker divided the process into four phases: AI Enabling Engineer, AI Enabling Staff, AI Enabling Team, and AI Enabling Organizations.
In the first phase, fast hands optimized the R & D paradigm through product upgrading, allowing AI to move from code-aid to Agent's autonomous R & D. In the second phase, fast-trackers extended the use of AI from engineers to all employees, allowing full access to AI for staff in different positions.
The third phase represents an important turning point in fast-track practice. Through a large body of practice, the technical teams were divided into four categories by fast-trackers: complex business development teams, AI-friendly development teams, CSTF teams and AI-based delivery teams.
The level of participation and autonomy of the four categories of teams in the AI R & D process varies, and AI evolves progressively from supporting individual R & D missions to undertaking the full R & D delivery process.
Among the non-technical teams, the quick-hander offers the “first distillation, then folding” approach: first, personal experience is amenable to personal competence through digital staff, reusable workflow, etc., and then individual experience is translated into the ability of a team to call together through knowledge, Skill and tools sharing.
Based on this idea, the MyFlicker 2.0 team AI product form was launched quickly. According to Shen Hwa, there are now more than 1,000 teams in the express handyard.
Phase four, fast start exploring the way from Team AI to Division AI. Shenlong pointed out that team AI was different in size and maturity, and that the whole organization was a completely new business, and that although its operating logic was similar to team AI, there were fundamental differences in the way it worked and in the integration of competencies.
In the case of the HR system, the organizational level AI needs to address not only individual tasks such as recruitment, performance, staff services, but also knowledge, data and business processes across functions.
To this end, fast-trackers are developing MyFlicker 3.0 to provide a stable Agent base for the organization of AI by increasing its workstation functionality, and to move AI from intra-team competencies to cross-team, cross-practice organizational collaboration.
Wang Dong Wuk: Super individuals, liquid tissues to skunk factories, quick to explore new models of organizational change
Wang Dong-hyuk, Director of Wind Control for Rapid Magnetic Engines, argued that the influence of AI on people presented a “smoke curve”: people with greater ability to use AI to further magnify their advantage, those with less capacity might be able to widen the gap and the Matthew effect became more pronounced.
In the face of this change, fast-trackers have explored three phases: from building “super individuals” to building “liquid tissues” to developing “skunk factory” models.
In the first stage, fast hands continually amplified individual capabilities through IDE, browsers Vibe Coding and CLI Agent. But super individuals do not automatically lead to increased team efficiency.
To this end, fast-tracking algorithms are integrated with engineering, front-end and back-end, allowing technicians with context operations to take on additional tasks and hand over some prompt projects to operators through technical alienation.
In the process of integration at the back and back, the quick-hands also built a multi-ware warehouse of Agent Harness. However, even with the full development of AI, there is still a ceiling for organizational efficiency.
Wang Dong Xue noted that the development chain accounted for approximately 20 per cent of the end-to-end delivery process and that overall delivery efficiency could only increase by about 22 per cent if AI increased code development efficiency 10 times.
As a result, the fast-tracker eventually chose the “elimination” of the organization's efficiency path, eliminating unnecessary coordination costs by binding common business objectives, reducing drift nodes, centralizing production research and decision-making, leaving small teams in a closed business circle, and gradually developing a “skunk factory” model.
According to Wang Dong Xu, as a result of this organizational and engineering adaptation, the team demand rate increased by 133 per cent, the delivery cycle decreased by 79 per cent, the total delivery demand increased by 800 per cent and the honor of the fast-track internal field-level AI template.
However, the skunk factory model also faces new challenges: external processes and cross-sectoral coordination may still pose obstacles when small teams need to work with large organizations.
Wang Dong-hyuk believes that breaking this bottleneck requires the emergence of more skunk factories and the gradual breaking of barriers to collaboration in large organizations through the collaboration of small teams with the capacity to deliver autonomously.
Hao Kong: Towards AI Native Organization, it needs to be top-down and bottom-up
The OPSO producer Hao has just shared the AI Native evolution path within OPPO: from universal access to tools, personal effectiveness to organizational effectiveness, to AI Native.
During the tool dissemination phase, OPPO introduced AI into day-to-day work such as programming, testing, operation, auditing, games and design, and even required staff to reach a monthly level of Token consumption. At the same time, OPPO, through AI README, has made project documentation and work elements more understandable to AI and has progressively built platform capacities such as digital human R & D, model squares and knowledge banks, and Skill and MCP centres.
As the AI tool became more widely available, OPPO found that an increase in individual efficiency did not mean a simultaneous increase in team delivery efficiency. To this end, OPPO has begun to increase the proportion of AI ' s participation and control in the R & D process, with a focus on two things: building a high-quality knowledge base and creating an AI delivery Z52Z that can evolve on a continuous basis to expand AI ' s synergistic delivery coverage and reduce overlap and waiting.
At the organizational level, the OPPO selected some of its personnel to conduct the AI transformation experiment, before gradually advancing the business process re-engineering, professional capacity deposition and platform base building. The Comind AI agent co-delivery platform, which is within it, is able to carry out a complete process from ideas, demand generation to end-line delivery, providing AI capacity support for team collaboration.
Hao has just argued that there is a need to move from top-down to bottom-up: managers need to take the lead and judge the direction of the transition; and employees need to move from super-individuals to synergetic “bees” that connect judgement, implementation and experience to each other, so that team profits are made.
II. AI is everywhere, and corporate AI needs to be credible
These issues pose new challenges when AI engages in in-depth team collaboration and enterprise R & D processes, how to stabilize Agent ' s mandate in complex business environments, how to measure the quality of delivery, and how to prevent new risks from efficiency gains. The two guests shared them at both methodological and practical levels, and how businesses should address them.
1 Zhu Xiaomin: the landing of enterprise-level AI agent requires a sophisticated engineering system
According to Zhu Min, author of Software Engineering 3.0, a professor-in-charge at Toji University, the landing of AI agent at the enterprise level does not focus solely on modelling capacity and the speed of code generation, but on the need for a reliable engineering system.
Enterprise productivity should be measured in terms of deliverable customer value, i.e. enterprise productivity = deliverable customer value/total input. If efficiency gains are accompanied by a decline in quality, deliverability and governance, the overall productivity of enterprises may decline.
He noted that generation was not equal to completion, personal performance was not equal to organizational efficiency, Demo was not equal to reliable production, and that enterprises needed to shift focus from generation to delivery, from individual efficiency to organizational throughput and from modelling capacity to engineering systems.
In the context of the concrete land-fall approach, Zhu focused on the statutory and managerial works. The statutory works were designed to define the objectives of Agent ' s work, context, constraints and acceptance check conditions. Harness Engineering establishes clear operational boundaries and operating environments for Agent through systematic engineering controls.
In addition, enterprises need to establish quality barriers. Given the probabilistic nature of Agent ' s behaviour, it cannot rely solely on its own judgement as to whether or not the task has been accomplished, but rather on the verification, interception and feedback of outputs through a stratification of quality, creating a closed circle of continuous improvement.
With regard to the enterprise ' s evolutionary path, Zhu suggested that three steps could be taken: first, structural clean-up and knowledge deposition to lay the foundation for Agent ' s involvement in research and development; secondly, the establishment of the Harness framework to create an engineering environment for Agent ' s autonomous mission; and finally, the construction of a quality door block and feedback loop that would allow Agent to move from reliance on external capacity and manual guidance to systemic autonomy.
2 Lu Winter and Winter: Build a system of testing of reasonable needs to enable AI to better test AI
According to Lu winter, the head of wind quality control in the fast-hand magnetic engine, AI is significantly increasing the speed at which individual soldiers are being delivered, but the increase in code generation does not mean that quality assurance capacity is growing simultaneously.
Validation capacity may become a new bottleneck when demand changes more frequently, code outputs become more fragmented, and even when there are hidden deficiencies that are superficially correct and actually semantic.
This change also poses new quality risks: testing does not keep pace with R & D and projects tend to accumulate at quality validation points; AI-generated test examples and conclusions lack sufficient credibility; and the proliferation of repeat codes, their corruption, also makes quality clearance more difficult.
In addition, the team ' s subjective perception of AI ' s effectiveness may be stronger than the actual output and can easily lead to overly optimistic judgements.
In response to these problems, Lu has shared the need-testing smart classification system that is being explored by a fast-track team. The team combines manual diagnostic experience with AI analytical capacity, establishes an intelligent hierarchical process that classifies demand into three high-, medium- and low-risk levels and matches different testing strategies: high-risk demand is artificially driven, supported by AI; medium-risk demand is AI and manually synergistic; low-risk demand is more AI-led.
The classification process also introduced a risk mining system consisting of 12 categories of detection items, covering such factors as core business chain changes, data security and compliance, context dependency and complexity, while retaining the highest manual decision-making permission.
As AI gradually takes on additional testing tasks, the quality team will also need to redefine the line of responsibility. Lu winter suggested that smart tests involve Agent gradually in determining “what to measure, how to measure, how to measure” depth, and that a combination of risk matching tests, based on a model of strategic decision-making, with an explanation of the strategy, eventually result in a structured test plan.
At the test implementation stage, multi-tool synergy, self-implementation, stability and downscaling are needed to move the test from manual to AI agent-driven.
However, the involvement of AI in the tests does not mean that it can be trusted directly by the conclusions given by AI. In particular, in the “AI-testing” scenario, the target itself is uncertain and the test tool's judgement may also be biased.
To this end, the fast-track team explored several ways to safeguard credibility: reducing single model deviations by cross-validation from different sources, testing test capabilities for mutations, calibrating AI by artificial real values, and insisting on quality decisions based on evidence rather than conclusions.
Conclusion: From individual effectiveness to organizational synergy, AI influences business collaboration patterns in depth
From the personal use of the AI writing code to the teamwork of the team through AI agent to the enterprise re-engineering of the organization and the R & D process, the changes brought about by AI are moving from a single-point tool to a re-engineering of the enterprise productivity system.
The introduction of AI by an enterprise ultimately depends not only on how fast codes are written and how efficient individual employees are, but also on the ability to make teamwork smoother and delivery more reliable and to sink personal experience into the ability of the organization to be used continuously.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.