Alibaba Cloud collapsed product and front-end work into PDFE

Alibaba Group employs over 100,000 workers, but accelerating individual programmers inside large teams has exposed a structural bottleneck. Data from Alibaba Cloud CIO Jiang Linquan's team shows that developers spend only about 20% of their working hours writing code, as reported by geekpark. The rest dissolves into coordination and handoffs. As Alibaba Group Chief People Officer Jiang Fang argued in geekpark, assembling a collection of AI-empowered super individuals does not naturally create a super organization.

Isolated speed thrives in solitary workflows. When single contributors use models to bridge skill gaps, work moves quickly, but the gains stall when code must pass between corporate silos.

Solitary developers already blur old boundaries. According to ifanr, Meng Lingya originally focused on back-end logic before using AI to handle front-end tasks on the Xiangji (响记) application. On Pixel Craft PRO (像素匠 PRO), Wang Xueyuan combined an astronomy academic background with front-end and product development. Deng Ke, born after 2000, built the HarmonyOS client for the sports app Xiaoqiuquan (小球圈) independently. Replicating that velocity across enterprise engineering demands collapsing functional handoffs into unified pods.

The review bottleneck behind solo developer speed

Solo contributors can now ship entire applications without relying on traditional functional teams. According to ifanr, Deng Ke, a developer born after 2000, developed the HarmonyOS client for the sports app Xiaoqiuquan (小球圈) independently. Other programmers cross structural specializations just as quickly. Developer Meng Lingya originally focused on back-end logic before using AI to handle front-end tasks on the Xiangji (响记) application. At Pixel Craft PRO (像素匠 PRO), developer Wang Xueyuan relied on an astronomy academic background while participating in front-end and product development without standard software credentials.

That individual speed fails to scale automatically across an enterprise. According to geekpark, Alibaba Group Chief People Officer Jiang Fang argues that assembling a collection of AI-empowered super individuals does not naturally create a super organization.

Inside Alibaba Group, which employs over 100,000 workers, isolated efficiency collides with team dependencies. Drafting code represents only a fraction of the working week. Data from Alibaba Cloud CIO Jiang Linquan's team shows that developers spend only about 20% of their working hours writing code. The remaining time goes toward verification and system alignment. According to data from Jiang Linquan's team, core and complex code accounts for 30% or less of total code but consumes 80% of senior developers' time. When solo creators churn out assisted drafts, senior engineers choke on review debt instead of building systems.

Collapsing handoffs into single-owner pods

Friction always accumulates at the handoff. According to geekpark, Alibaba Cloud CIO Jiang Linquan's team tackled this lag by creating a unified role called PDFE. This position absorbed product management and interface design directly into front-end development duties. Back-end engineering and systems architecture merged into a counterpart role named ABE. API contracts now connect them. That technical boundary replaced negotiation cycles over static mockups.

Fusing these duties strips out coordination overhead. A developer moves straight from requirements to functional client code, governed entirely by the data schema.

This structural consolidation produced immediate velocity gains. Geekpark reported that Alibaba's Qoder team developed QoderWork with a five-person team in seven days. The broader division soon followed. In February, Alibaba Cloud's public cloud business unit launched a mandatory 100-day AI maker competition for all of its employees. The initiative tested whether small pods could sustain that pace across production workflows.

Single-owner execution relies on packaging complex technical layers into accessible entry points. According to ifanr, HarmonyOS DevEco CLI bundles platform toolsets and official knowledge bases alongside Skills into command-line interfaces. Agents take over from there. They use the CLI to manage code generation, compilation, deployment, and debugging. HarmonyOS 7 takes a matching approach with FAST Kit, packaging low-level algorithms like signal processing and natural language sorting into ready-to-use services. The Xiangji (响记) application demonstrates this pattern. As ifanr reported, it calls the HMS_FAST_Algo_NaturalSort interface from FAST Kit to sort filenames in natural language sequence.

Relocating agents from private chats to shared rooms

Shared project spaces pull automated agents out of private windows and into open view. In July, Jack Dorsey's company Block launched Buzz, an open-source collaboration project that integrates AI agents such as Claude Code and Codex into communication channels. Coze 3.0 operates on an equivalent premise. The platform allows multiple human users and multiple AI agents to collaborate in a single project, including external agents like Claude Code and Codex. Kylon organizes its workspace around Rooms that house human team members, AI agents, files, and custom internal applications. To preserve team context, Kylon manages AI memory by compressing long chat histories into searchable summary trees.

Shared interfaces alone cannot dissolve institutional boundaries. Moving assistants into group channels exposes private queries. Still, visibility does not create operational alignment.

According to woshipm, departments at a leading Chinese e-commerce company began passing AI-generated proposals back and forth, consuming company-funded tokens without increasing organizational efficiency. Work expanded without moving forward. Mu Sheng argues that AI enhances individual worker efficiency, but individual gains cannot translate into organizational efficiency without restructuring the organizational model. Enterprise AI adoption without organizational change reinforces departmental silos because different business units deploy separate AI models to justify their positions. Confronting this friction, Alibaba Group Chief People Officer Jiang Fang and Alibaba Cloud Chief Information Officer Jiang Linquan jointly spoke about organizational transformation at the Apsara Conference AI Native Organization Forum, according to geekpark.

Restructuring Enterprise Workflows from Isolated Prompts to Collaborative Pods

• Cross-departmental teams generate separate AI proposals back and forth, consuming company-funded tokens while entrenching departmental silos and stalling delivery. Shift from disconnected prompt tools to shared multi-agent workspaces like Kylon's Rooms or Coze 3.0-which support external agents like Claude Code and Codex and condense long histories into searchable summary trees-to unify human and agent context in a single workspace.
• Developers spend only about 20% of their working hours writing code, while complex code represents 30% or less of total code but consumes 80% of senior developers' time due to functional handoffs. Collapse traditional functional boundaries by adopting Alibaba Cloud's pod structure: combine product management, design, and front-end development into a single PDFE role, interfacing with an architecture and back-end (ABE) role strictly through API contracts.
• Management evaluates AI adoption strictly through individual contributor output, per-capita revenue metrics, or workforce reductions rather than end-to-end business velocity. Implement Mu Sheng Consulting's organizational metrics: track token efficiency (financial or operational output divided by total labor plus AI expenditure) and audit cycle duration in days across three core decision workflows: budget additions, cross-departmental resource reallocations, and new product launches.
• Small product teams or domain specialists need to build, test, and ship full-stack applications without waiting for specialized infrastructure or front-end departments. Equip small cross-functional pods with integrated agent command-line toolchains like HarmonyOS DevEco CLI and FAST Kit-mirroring Alibaba's five-person Qoder team that delivered QoderWork in seven days, or solo developers using Agentic Coding to handle full-stack execution.

System delegation and the cost of enterprise integration

Scaling automated workers requires infrastructure overhauls rather than isolated chat boxes. According to geekpark, general agent invocation volume increased by 1,000 times over a four-month period after Alibaba Cloud redesigned its systems for general AI agents. Alibaba Cloud deployed 28 types of digital employees, equivalent to over 2,000 employee headcount. Overhead fell quickly. Its translation tool Ruiyibao (睿译宝) achieves a 95% simultaneous interpretation accuracy rate at one-fifth to one-tenth the cost of human labor. Meanwhile, the outbound calling tool Ruihubao (睿呼宝) operates at one-fifth the cost of human agents while matching human performance in renewal reminders and trial-to-paid conversion. Lead cleaning shows identical parity.

System delegation works only when agent activity ties directly to financial output. Raw token volume reveals very little about organizational health.

According to woshipm, Mu Sheng Consulting defines token efficiency as financial or operational output divided by total human and machine costs, specifically labor costs plus total AI expenditure. Impact appears in decision velocity. Mu Sheng Consulting evaluates management speed by tracking the duration in days for budget additions alongside cross-departmental resource reallocations. It applies that measurement to new product launches as well. Mu Sheng argues that consolidating TRAE, Coze (扣子), and Doubao into Doubao Work (豆包工作) turns ByteDance's AI workplace tooling from an internal cost center into an external-facing operating unit accountable for business results.

Connecting these tools into enterprise systems demands shared interaction protocols. According to ifanr, Xiaoqiuquan connects with Huawei's Celia (Xiaoyi) assistant via HarmonyOS's system-level Agent-to-Agent framework to support voice queries for sports statistics. Protocols create consistency. As woshipm notes, Kylon supports integrations with more than 3,000 third-party services, including Notion and GitHub. When digital workers run through governed application programming interfaces instead of arbitrary text windows, companies protect institutional context while maintaining operational control.

Organizational and Architectural Approaches to Multi-Agent Workspaces

Engineering groups should evaluate structural boundaries instead of monitoring individual prompt speed. According to geekpark, Alibaba Cloud CIO Jiang Linquan's team merged product manager and design duties directly into front-end development, establishing the single PDFE role. Back-end development and architecture combined into ABE, linking to front-end work across API contracts.

That structural rewrite cleared friction immediately. Alibaba's Qoder team assembled QoderWork with a five-person team in seven days.

Scale follows structural consolidation rather than isolated chat windows. Over a four-month period after Alibaba Cloud redesigned its systems for general AI agents, general agent invocation volume rose by 1,000 times. Engineering leaders can measure how shared workspace infrastructure lets smaller groups ship finished software without functional handoffs.

For readers outside China

• Availability: Block's Buzz was launched in July as an open-source collaboration project. Kylon, backed by seed funding from Leonis Capital, supports integrations with global platforms such as Gmail, Notion, and GitHub, though its regional deployment limitations are not disclosed in sources. ByteDance announced the integration of TRAE and Coze into Doubao to launch Doubao Work on August 24, but international corporate access is not disclosed in sources. Huawei's HarmonyOS 7 developer kits and Alibaba Cloud's internal operational tools (such as Ruiyibao and Ruihubao) are embedded within their proprietary platforms, and external availability outside China is not disclosed in sources.
• Pricing: Base subscription fees and token pricing models are not disclosed in sources for Kylon, Coze 3.0, TRAE, Doubao Work, or HarmonyOS DevEco CLI. Sourced operational costs are comparative: Alibaba Cloud's translation tool Ruiyibao runs at one-fifth to one-tenth the cost of human labor, while its automated outbound calling tool Ruihubao operates at one-fifth the cost of human agents.
• Closest Western equivalents: Block Buzz; Claude Code; OpenAI Codex
• Data residency: The source material does not cover data residency protocols, server locations, or cross-border data transfer policies. Workflows built on ByteDance foundation models (such as Skylark, Seedance, and Seedream) or Huawei's Celia assistant and Agent-to-Agent interaction framework run within domestic Chinese infrastructure, but specific compliance details for foreign corporate data are not disclosed in sources.

Sources

• geekpark 从超级个体到超级组织，究竟还有多远？ https://geekpark.net/news/371104
• ifanr 小团队的大创意，为什么能在 HarmonyOS 7 上变为现实？ https://ifanr.com/1682085
• woshipm Manus肖弘昔日大学搭档创业，Kylon做了个AI原生工作空间 https://woshipm.com/ai/6468012.html
• woshipm 你的HR还在算人均营收？字节已把AI变经营单元！ https://woshipm.com/ai/6453979.html