Introduction
The landscape of AI agent products has undergone dramatic shifts between China and the United States. Early competition centered on coding‑specialized agent capabilities. As the industry evolves, the competitive focus is expanding toward full‑fledged workplace agent platforms that handle daily office workflows. Tencent WorkBuddy’s market launch marks one critical milestone. Within two‑week windows, major domestic vendors rolled out organizational adjustments and product iterations. Meanwhile, overseas players such as Anthropic keep advancing their office‑oriented agent roadmap. Divergent development paths, business models and data flywheel mechanisms have created measurable gaps between Chinese and US AI agent ecosystems. This article sorts out market metrics, corporate strategic choices, root‑cause disparities, and the new competitive phase built around general‑purpose Work agents. When enterprises operate multi‑vendor agent traffic with mixed coding and office workloads, teams may leverage an API gateway such as 4sapi to unify access and traffic governance for heterogeneous LLM endpoints.
Tencent WorkBuddy Launch and Recent Industry Restructuring
Tencent officially released WorkBuddy in March 2026. After more than three months of public testing, third‑party analytical statistics show that WorkBuddy recorded 20.97 million monthly visits on China’s PC‑end AI native office agent market, securing first‑place ranking. Its monthly active figure was nearly 2.4 times higher compared with the 8.85 million baseline set back in March release. It is worth noting that office‑oriented products generate large user volume while coding‑related businesses deliver higher short‑term revenue.
A wave of organizational reshuffles swept major Chinese AI vendors within a short two‑week window.
- July 20: Tencent consolidated QClaw teams into its product division six, placing them under the same management framework as WorkBuddy.
- July 30: ByteDance integrated its self‑developed agent products including CodeWorker, MuleRun and Wukong, appointing new core leadership for the new Work‑agent‑focused product unit.
- August 30: Moonshot AI launched Kimi Work, evolved from its prior coding‑focused Kimi Code.
These synchronized adjustments are not pure coincidence. On the global track, Anthropic launched Claude Code earlier and later rolled out Claude Cowork targeting ordinary office knowledge‑workers. According to internal industry calculations from Moonshot AI, coding scenarios consume over 90 % of total token volume, yet office‑knowledge‑worker‑oriented scenarios hold more than one billion potential users. Many insiders regard the coming two‑to‑three‑year window as a critical inflection point for agent‑driven paradigm shifts.
WorkBuddy inherits technical assets accumulated from CodeBuddy, Tencent’s coding agent developed three years earlier. After two years of iterative refinement, engineering teams built a generic Agent kernel embedded inside WorkBuddy. Product leaders acknowledge realistic gaps against top‑tier overseas models: domestic‑only model stacks cannot fully match capabilities demonstrated by Claude Opus. Even so, starting from mid‑2025, product‑level gaps on coding tasks have narrowed substantially in benchmark testing.
From business‑strategy perspective, Chinese vendors are shifting priority from pure coding agents toward comprehensive Work agents. Multiple driving factors support this transition. First, office‑oriented scenarios bring massive user groups covering administration, sales, human‑resources and operational roles beyond developer circles. Second, user stickiness is extremely high: once users adopt agent‑assisted office workflows, switching costs become substantial. Third, business value multiplies. While coding‑agent consumption sits at individual‑developer scale, Work‑agent workloads can expand usage magnitude by 100‑fold or even 1000‑fold once adopted across entire enterprise teams.
Challenges remain for office‑oriented agent adoption. User‑uploaded local files, diverse enterprise internal documents and complex permission boundaries raise higher requirements for data security and compliance. Monetization paths are still being explored. One envisioned direction is knowledge‑base‑linked metered billing. Industry practitioners point out that Silicon Valley players including Codex and ChatGPT are also moving toward unified workplace‑entry‑point positioning, where AI assistance extends beyond working hours.
Widened Gap: Asynchronous Development of Sino‑US Coding‑Agent Ecosystems
Back in October 2024, US coding‑agent startup Bolt.new went online. Within four weeks, its annualized revenue hit 4 million US dollars. Anthropic executives publicly noted Bolt.new became their fastest‑growing customer, whose surging API calls saturated Anthropic GPU resources. This case illustrates how rapidly developer‑oriented agent businesses could scale in the United States.
Statistical data reveals uneven market distribution. As of early August 2026, developer‑oriented agent products built upon Anthropic services consume 24.9 % of total token usage but capture 71.8 % of total market spending. Cursor alone has exceeded 400 million US dollars in annualized revenue. By contrast, none of GitHub, Tencent or ByteDance have released comparable public revenue figures for their coding‑agent businesses.
Mid‑2025 marks an important dividing line for industry trajectories. Before that point, most Chinese large‑model companies primarily benchmarked OpenAI’s feature set, replicating capabilities such as universal dialogue, image and video generation. Anthropic took a different technical path. As early as 2021, its engineering teams began exploring code‑assistant prototypes. Around early 2022, internal teams built tool‑calling pipelines for autonomous software‑engineering tasks. By 2023‑2024, the internal prototype cilde had already taken shape, which later evolved into Claude Code.
The core difference lies in data flywheel construction. Anthropic regarded coding‑agent products not merely as capability‑output channels, but also as critical data‑ingestion entry points. Real‑world software‑engineering workflow data — prompt inputs, code modification attempts, test failures, evaluation feedback — continuously flows back to improve base models. More users bring higher‑quality real‑world code datasets; better models attract more developer users, forming self‑reinforcing positive cycles.
Chinese enterprises also identified commercial opportunities within coding scenarios. Traditional finance, manufacturing and internet sectors generate real demand for code‑assistant tools. However, many domestic implementations center on customized project delivery and on‑site deployment patterns. This project‑based sales model can secure short‑term revenue, yet struggles to produce exponential user‑growth flywheels.
Overseas independent coding‑agent startups adopted standardized SaaS and MaaS patterns. Products such as Lovable, Bolt.new and GitHub Copilot deliver annual‑revenue figures measured in hundreds of millions US dollars. Their business logic focuses on increasing conversion rates and higher‑value API consumption instead of one‑off project contracting.
Talent background also shaped product philosophy. When Boris Cherny joined Anthropic in September 2024, internal experiments proved agents should not be limited to responding passively to human prompts. Agents could receive high‑level objectives, then autonomously read, write, execute programs and iterate based on runtime feedback. This idea became one foundational design principle for Claude‑series agent products.
Turning‑point: Chinese model vendors accelerate catch‑up driven by new competitive variables
The release of DeepSeek R1 in early 2026 delivered strong external stimulus for the whole Chinese agent industry. Multiple leading domestic firms readjusted resource allocation toward coding and agent tracks.
- Zhipu AI reallocated engineering manpower, releasing GLM‑4.5 and advancing coding, agent and reasoning capabilities.
- Moonshot AI slowed large‑scale consumer‑oriented marketing campaigns and refocused resources on underlying model iteration. Kimi Composer quickly gained community traction after launch. It is one representative case where Chinese open‑weight models power end‑user agent workflows at scale.
Nevertheless, one structural risk persists. If vendors only publish open‑weight model checkpoints while third‑party developers add harness layers and user feedback loops, the most valuable real‑world interaction datasets will accumulate on third‑party application sides rather than original model providers. This reality pushes domestic foundational‑model companies to build their own front‑end agent products.
Major Chinese cloud and AI giants ramped up internal investment:
- Alibaba assembled dedicated internal teams for its Qoder agent product.
- Tencent kept iterating CodeBuddy and WorkBuddy.
- ByteDance sustained investment on the Trae coding agent stack and corresponding enterprise offerings.
Traditional assumptions stated only well‑resourced big corporations could afford foundational‑model training. Circumstances shifted by 2026. Model startups can concentrate computing resources targeting high‑priority vertical agent scenarios. Large enterprises meanwhile adopt multi‑model hybrid strategies: combining self‑developed base models, open‑weight third‑party models, cloud infrastructure and enterprise sales channels to seek commercial returns from coding‑agent businesses.
Competition definition has evolved. Two years ago, Chinese players mainly chased pure coding‑agent capability. Now industry consensus forms: catching up on coding benchmarks is feasible given high‑quality engineering talent pools and abundant application‑scenario resources. But the competitive battlefield expands further. The race is moving from pure coding agents toward comprehensive Work‑agent platforms covering document reading, tool invocation, program execution and multi‑hour autonomous iterative workflows.
Conclusion
The earlier round of global coding‑agent competition saw Chinese participants fall behind in product launch timing, revenue realization, model iteration and data‑flywheel construction. Now the industry enters a brand‑new Work‑agent competition phase. Business models, compliance frameworks and product‑market fit for universal workplace agents remain unproven.
US players built advantages from mature developer‑oriented SaaS ecosystems and well‑established agent‑data feedback loops. Chinese competitors hold strengths including massive domestic enterprise‑office user bases, rich real‑world document scenarios and aggressive resource input from both large corporations and model startups.
The competition is no longer purely about abstract benchmark scores. It revolves around end‑to‑end product experience, secure data governance, sustainable monetization and closed‑loop data collection. Whether Chinese vendors can realize competitive overtaking hinges on how well they construct positive cycles among product adoption, real‑world workflow data and continuous model improvement.
