Kuaishou proposes the "Skunk Works" model: How AI moves from individual productivity gains to organizational transformation

Zhidongxi reported on October 31 that today, at a technology salon, Kuaishou, drawing on the practices of its enterprise-level AI productivity platform iWork, shared how AI is gradually moving from assisted programming toward team collaboration and organization-level efficiency gains, and is being consolidated into enterprise productivity infrastructure.

Shen Lang, head of Kuaishou's enterprise agent development and an enterprise AI productivity expert, reviewed the four stages of Kuaishou's AI productivity development over the past two years: from empowering engineers with AI, to empowering all employees, to empowering teams and the organization with AI. He particularly noted that individual efficiency gains do not equal team efficiency gains; enterprises need to consolidate individual experience into reusable digital employees, workflows, and Skills, further promoting team capability sharing and organization-level Agent collaboration.

Zhu Shaomin, author of "Software Engineering 3.0" and a distinguished professor at Tongji University, started from engineering methodology and shared the key path for implementing enterprise-level agents, emphasizing the establishment of an engineering system for reliable Agent delivery through specification engineering, harness engineering, and quality gates.

Wang Dongxu, head of risk control technology at Kuaishou Magnetic Engine, shared from the perspective of organizational change Kuaishou's exploration of moving from "super individuals" to "liquid organizations" and "skunk works," attempting to break down job boundaries and reduce collaboration steps so that the technical efficiency brought by AI can truly translate into organizational efficiency.

Hao Gang, head of product for R&D efficiency at OPPO, shared OPPO's internal AI Native evolution path, introducing practices from tool popularization and individual efficiency gains to organizational efficiency gains, and how to promote the integration of AI into the complete R&D process through high-quality knowledge bases, AI delivery workflows, and the agent collaborative delivery platform CoMind.

Facing challenges such as frequent requirement changes and insufficient credibility of AI-generated test cases, the team is exploring intelligent grading of requirement testing, risk mining, and autonomous agent testing, and is addressing the credibility issue in the "AI testing AI" process through methods such as heterogeneous cross-validation and manual ground-truth calibration.

As AI tools gradually become widespread, the problem enterprises face is how to transform the efficiency gains scattered among individual employees into team collaboration and organizational capability. Three guests from Kuaishou and OPPO shared different paths for enterprise AI transformation from individual efficiency gains to organizational change.

1. Shen Lang: From individual to organization, the four-stage evolution of Kuaishou's AI productivity

Shen Lang, head of Kuaishou's enterprise agent development and an enterprise AI productivity expert, believes that enterprise AI transformation needs to go through a gradual evolution from individual to team and then to organization. Kuaishou divides this process into four stages: AI empowering engineers, AI empowering all employees, AI empowering teams, and AI empowering the organization.

In the first stage, Kuaishou mainly optimizes the R&D paradigm through product upgrades, allowing AI to gradually move from code assistance to Agent autonomous development. In the second stage, Kuaishou expands the scope of AI use from engineers to all employees, enabling employees in different roles to fully use AI.

In non-technical teams, Kuaishou proposed an efficiency improvement approach of "distill first, then fold": first AI-ify individual experience, consolidating individual capabilities through digital employees and reusable workflows, and then, through sharing knowledge, Skills, and tools, transform individual experience into capabilities that the team can jointly invoke.

Based on this approach, Kuaishou launched the MyFlicker 2.0 team AI product form. According to Shen Lang, more than 1,000 team Agents are currently running in Kuaishou's campus.

In the fourth stage, Kuaishou began exploring the move from team AI to organizational AI. Shen Lang pointed out that team AI varies in scale and maturity, while the entire organization is a brand-new business format; although its operating logic is similar to team AI, there are essential differences in collaboration methods and capability integration.

Taking the HR system as an example, organizational-level AI not only needs to handle individual tasks such as recruitment, performance, and employee services, but also needs to connect knowledge, data, and business processes across different functions.

To this end, Kuaishou is developing MyFlicker 3.0, adding workbench functionality to provide a stable Agent foundation for organizational AI, promoting AI's further move from capability reuse within teams to cross-team and cross-business organizational collaboration.

2. Wang Dongxu: Super individuals, liquid organizations, and then skunk works, Kuaishou explores a new model of organizational change

Facing this change, Kuaishou explored three stages: from creating "super individuals," to building "liquid organizations," and then to forming the "skunk works" model.

In the first stage, Kuaishou continuously amplified individual capabilities through methods such as IDE, browser Vibe Coding, and CLI Agent. But super individuals do not automatically bring team efficiency improvements.

To this end, Kuaishou promoted the integration of algorithms and engineering, and of frontend and backend, allowing technical personnel with business context to take on more work, and also, through technology transfer, handing over part of prompt engineering to operations personnel.

In the frontend-backend integration process, Kuaishou also built a multi-repository full-stack Agent Harness. However, even with AI full-stack development, there is still a ceiling for organizational efficiency.

Wang Dongxu pointed out that the development phase accounts for about 20% of the end-to-end delivery process; even if AI increases code development efficiency by 10 times, overall delivery efficiency can only improve by about 22%.

Therefore, Kuaishou ultimately chose the organizational efficiency improvement path of "elimination," eliminating unnecessary coordination costs by binding common business goals, reducing handoff nodes, and centralizing product, operations, and R&D decision-making, allowing small teams to form business closed loops and gradually forming the "skunk works" model.

According to Wang Dongxu, after this adjustment of organization and engineering methods, the team's requirement throughput increased by 133%, delivery cycle was shortened by 79%, the number of full-stack delivery requirements increased by 800%, and it won the honor of Kuaishou's internal domain-level AI model room.

However, the skunk works model also faces new challenges: when small teams need to collaborate with large organizations, external processes and cross-department coordination may still create obstacles.

Wang Dongxu believes that breaking this bottleneck requires the emergence of more skunk works, and through the collaboration of multiple small teams with independent delivery capabilities, gradually breaking down collaboration barriers in large organizations.

3. Hao Gang: Moving toward an AI Native organization requires both top-down and bottom-up approaches

Hao Gang, head of OPPO's R&D efficiency improvement products, shared OPPO's internal AI Native evolution path: from tool popularization and individual efficiency improvement to organizational efficiency improvement, and ultimately toward an AI Native organization.

In the tool popularization stage, OPPO introduced AI into daily work such as programming, testing, operations, review, gaming, and design, and even once required employees to reach a certain monthly Token consumption level. At the same time, through AI README, OPPO makes project documentation and work content easier for AI to understand, and gradually builds platform capabilities such as digital human R&D, model marketplace and knowledge base, as well as Skill and MCP centers.

As AI tools gradually became widespread, OPPO also found that individual efficiency improvement does not mean team delivery efficiency improves in tandem. To this end, OPPO began to increase the proportion of R&D processes in which AI can participate and take control, and focused on advancing two things: building a high-quality knowledge base, and creating a continuously evolving AI delivery workflow to expand the coverage of AI collaborative delivery and reduce repetitive handoffs and waiting.

At the organizational level, OPPO first selected some personnel to carry out AI transformation experiments, and then gradually advanced business process restructuring, accumulation of professional capabilities, and platform foundation construction. Its internal CoMind agent collaborative delivery platform can undertake the complete process from ideas and requirement proposals to final online delivery, providing AI capability support for team collaboration.

On how to move toward AI Native, Hao Gang believes that both top-down and bottom-up paths need to be advanced together: managers must not only get hands-on themselves but also judge the direction of transformation; employees need to move from super individuals to collaborative "swarms," connecting judgment, execution, and experience with one another to form team compounding.

As AI penetrates team collaboration and enterprise R&D processes, how to enable Agents to stably execute tasks in complex business environments, how to measure delivery quality, and how to prevent efficiency improvements from bringing new risks have become new challenges. The two guests shared, from the levels of methodology and practice respectively, how enterprises should solve these problems.

1. Zhu Shaomin: Enterprise-level agent implementation requires a sophisticated engineering system

English text is being prepared.

Enterprise productivity should be measured by deliverable customer value, that is, enterprise productivity = deliverable customer value / total investment. If efficiency improvements are accompanied by declines in quality, deliverability, and governability, overall enterprise productivity may instead decrease.

He pointed out that generation does not equal completion, individual efficiency improvement does not equal organizational efficiency improvement, and a Demo does not equal reliable production. Enterprises need to shift their focus from generation speed to delivery speed, from individual efficiency to organizational throughput, and from model capabilities to engineering systems.

In terms of specific implementation methods, Zhu Shaomin focused on specification engineering and harness engineering. Specification engineering aims to define the Agent's work objectives, context, constraints, and acceptance conditions. Harness Engineering, through systematic engineering control, establishes clear execution boundaries and operating environments for Agents.

In addition, enterprises also need to establish quality gates. Because Agent behavior is probabilistic, one cannot rely solely on the Agent itself to judge whether a task is complete; instead, through layered quality checks, outputs should be verified, intercepted, and fed back to form a closed loop of continuous improvement.

Regarding the evolution path for enterprises, Zhu Shaomin proposed that it can be advanced in three steps: first, carry out architecture cleanup and knowledge accumulation to lay a foundation for Agents to participate in R&D; second, build a Harness framework to establish an engineering environment for Agents to autonomously execute tasks; finally, build quality gates and a feedback loop so that Agents can gradually move from relying on external capabilities and manual guidance to system autonomy.

When requirement changes become more frequent, code output becomes more fragmented, and even some hidden defects appear that are superficially correct but semantically wrong, verification capability may become the new bottleneck.

This change also brings new quality risks: testing speed cannot keep up with the R&D pace, and projects tend to pile up at the quality verification stage; AI-generated test cases and conclusions lack sufficient credibility; the increase in duplicate code and accelerated code decay also make quality review more difficult.

In response to these problems, Lü Dongdong shared the requirement testing intelligent grading system that the Kuaishou team is exploring. His team combines manual judgment experience with AI analysis capabilities to establish an intelligent grading process, dividing requirements into three risk levels—high, medium, and low—and matching them with different testing strategies: high-risk requirements are led by humans with AI assistance; medium-risk requirements are handled collaboratively by AI and humans; low-risk requirements are more often led by AI.

In the grading process, a risk mining system containing 12 categories of detection items was also introduced, covering factors such as changes to core business links, data security and compliance, context dependencies, and complexity, while retaining the highest manual decision-making authority.

As AI gradually takes on more testing work, quality teams also need to redefine responsibility boundaries. Lü Dongdong proposed that intelligent testing should allow Agents to gradually participate in deciding "what to test, how to test, and how deeply to test," using a strategy decision model to match test combinations according to risk and provide explainable strategic basis, ultimately forming a structured test plan.

In the test execution stage, it is necessary to further achieve multi-tool collaboration, autonomous execution, stability and reliability, and cost reduction and efficiency improvement, so that testing can gradually move from manual operation to agent-driven.