AI Coding · OpenAI

10,000 Chinese Codex Clones: Desktop Agents Fight for the Execution Layer

By 莪_幻尘 · Oct 23, 2026

Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Whoever controls the desktop execution layer becomes the default way people tell computers what to do, and that changes the competitive field from chat answers to delivered results. For developers outside China, the same shift is coming to agent tooling and enterprise software, so the paths and failure modes on display here are a preview of choices coming to their own stacks.

Summary

Codex's pivot from CLI coding tool to desktop agent changed the stakes: by June 3, 2026 it had 5 million weekly active users, up more than 7x in six months, and most users were not engineers. Chinese vendors responded fast. Alibaba, Tencent, ByteDance, Baidu, Moonshot, MiniMax, Zhipu, StepFun and NetEase all shipped desktop agents between January 30 and June 24, 2026, with Tencent alone releasing three.

These 15 products do not converge on one form. Some carry coding-agent strengths into office tasks (Kimi Work, MiniMax Code), some sit close to the OS and operate files, screens and apps directly (Marvis, QoderWork, Cola), and some embed into existing ecosystems like Feishu or WeCom (TRAE Work, WorkBuddy). The underlying trends are the same: AI is moving from answering questions to delivering results, from a single assistant to agent teams, and from inside applications up to an intent execution layer over the OS.

Takeaways

— Codex reached 5 million weekly active users by June 3, 2026, up more than 7x in six months, and non-engineers made up most of its users.

— At least 15 desktop agents launched in China between January 30 and June 24, 2026, with Tencent shipping WorkBuddy, Marvis and Qclaw.

— Alibaba's QoderWork launched first and added memory, reflection and skill-evolution "consciousness features" on June 16.

— ByteDance upgraded TRAE SOLO into TRAE Work, which deeply integrates Feishu and offers Work and Code modes, and also launched Doubao Professional.

— Kimi Work can autonomously create sub-agent teams and scale one capability cluster to 300 collaboration units.

— WorkBuddy bundles 11 domestic models in Auto or manual mode, with token consumption varying by more than 20x between models.

— Doubao Professional's task mode is free for non-paying users with the 2.1 Turbo model, a rarity in a paid-suite market.

— Qclaw keeps data local and Niuma AI is fully offline, while ecosystem-embedded products like WorkBuddy and TRAE Work necessarily touch cloud documents and collaboration data.

— Three competitive paths emerged: engineering spillover from coding agents, OS-level control of files and apps, and embedding into existing ecosystems.

— Three trends define the wave: AI coding expands into AI working, single agents expand into agent teams, and the battle moves from inside apps to an intent execution layer above the OS.

Conclusions

Codex's user base being mostly non-engineers is the reason Chinese vendors are building office assistants and OS-level agents rather than better coding CLIs; the product line follows the user mix.

The split into three paths, instead of one winning form, means no one has yet proven the right abstraction for desktop agents; each path has a different bottleneck, from task decomposition to OS permissions to ecosystem lock-in.

Agent teams shift the user's job from operating a tool to managing a small workforce, which only pays off if the underlying engineering around context, rollback and approval is mature; otherwise multiple agents just add noise.

Concepts & terms

Desktop agent

A software agent that works inside a user's desktop environment, directly handling files, browsers and applications to complete a task, instead of only returning an answer in a chat window.

Agent Team

A collaboration pattern where multiple specialized agents with different roles work in parallel on one task, coordinated by a mechanism that merges their output; MiniMax Code and Kimi Work use it for long, multi-stage work.

Intent execution layer

A software layer above the operating system that interprets a user's goal and operates files, apps, browsers and schedules on their behalf. It does not replace the OS kernel, but it can turn the OS into a pipe if users make it their default way of commanding a computer.

From the discussion

Featured comments

咬代码的兽 1 likes

The assessment of the execution entry point is spot on. What desktop agents are really fighting over is ownership of the user's default actions. For developers, multi-platform adaptation costs will keep rising, and cross-platform solutions are worth planning ahead for.

See top comments, translated →

Source: juejin.cn ↗ Google Translate ↗ Backup ↗