Over the past month, Tencent, Alibaba and ByteDance have all started doing the same thing: cutting back on AI agents instead of launching more of them.
Tencent issued an internal notice that moved all business under the Tencent QClaw product center, along with part of the team, into WorkBuddy. The piece says WorkBuddy has been described as a business with the potential to become Tencent’s third breakout product after QQ and WeChat.
Alibaba has also moved to consolidate. According to information obtained by Caijing, the company is preparing to launch “Qwen Office,” merging three agent products — QoderWork, Wukong and MuleRun — under the leadership of newly appointed DingTalk CEO Chen Yusen. The article says those three products were only handed to him in early July and have now been rolled into one. By placing its office-facing AI products under the Qwen brand, Alibaba’s intent is laid out plainly in the original text.
ByteDance has made a quieter change. Its AI coding product TRAE SOLO has been renamed TRAE Work. The article reads that one-word revision as a major shift in direction.
Any one of these would look like a standard organizational adjustment on its own. Taken together, with all three companies pulling a stack of agents launched over the past six months back toward a single entry point at nearly the same time, the article says it is the first real product consensus of the agent era.
Fewer agents, not more
The article looks back at the early burst of agent development this year, around the Lunar New Year period, when companies broadly believed agents could become the next mass platform. Teams built one, departments built one, and individual use cases got their own versions. In just a few months, thousands of agents with similar functions appeared.
Tencent is presented as a clear example. QClaw was built by the PC Manager team on top of OpenClaw, while WorkBuddy came out of Tencent Cloud. There were also QQ Longxia and Browser Longxia, with different versions spread across business groups and developed separately.
Alibaba, the article says, was also running several internal projects at once. These included desktop AI agent tool QoderWork, the still-evolving Wukong, and MuleRun, which targeted overseas markets. Their positioning was close, functions overlapped heavily, and both computing power and R&D resources were split. Users, in the article’s account, were left without a clear unified understanding of what the company was offering.
ByteDance also launched multiple agent products in succession, including ArkClaw, ByteClaw, Feishuaily and TRAE SOLO.
The same pattern appeared outside China. The article points to OpenAI, Anthropic and Google, all of which rolled out product lines including Operator, Deep Research, Canvas, Projects, Codex, Claude Code, NotebookLM and Gemini.
This was not an unusual stage for a new category. The piece says that during exploration, the best management approach is often to allow repetition. But experimentation costs money. Over the past six months, big tech firms ran several agent tracks in parallel and consumed large amounts of inference compute. Multi-step reasoning, repeated API calls and redundant context retrieval pushed per-run costs well above those of ordinary chat models. Enterprise monetization did not keep pace, and the fragmented product structure dragged down overall return on investment.
Once open-source tools flattened technical barriers, the nature of competition changed. Compute budgets could no longer support unlimited internal duplication, and resources had to be concentrated. In the article’s framing, exploration ends when the direction becomes clear.
Big tech is choosing subtraction
The article places this shift in a longer internet history. In the PC era, browsers unified the web. In the mobile era, super apps unified services. In the AI era, it argues, a “super workbench” is starting to unify agents.
It draws on earlier examples of product convergence. Tencent built WeChat, QQ, Qzone, Pengyou and Tencent Weibo, yet the mobile internet eventually converged around WeChat. Meituan spread across group buying, film, food delivery, travel and ride services, before becoming a super app. Didi developed fast rides, premium rides, carpooling and designated driving, then packed those services into a single Didi entry point.
In that reading, product convergence does not mean innovation failed. It means the market is moving toward certainty. A market matures not when the number of products keeps rising, but when the number starts to fall. Platform wars are ultimately decided not by how much a company can create, but by how much it can simplify.
The article then maps that logic onto the current round of changes. Tencent has taken experimental work scattered inside the PC Manager team and pulled it back into CSIG, where cloud services are the main focus, placing it inside WorkBuddy under Tencent Cloud’s control. Alibaba is giving up a decentralized model in which different business lines incubate their own AI assistants and is instead putting office AI under a DingTalk-led “Qwen Office,” while unifying the branding under Qwen. ByteDance has shifted TRAE toward workflow collaboration. The article says that move effectively ends the phase in which an agent stands alone as an independent unit, with SOLO turning into an invisible technical base inside TRAE Work.
In the history of the internet, the article says, each round of product unification marks the point when the real competition begins.
The market is larger than programmers
The article argues that a deeper turn is hidden beneath this convergence: programmers are no longer seen as the biggest market for AI.
Over the past year, many companies treated developers as the most important user base. Coding was one of the earliest AI scenarios to gain traction, and products such as Cursor, Claude Code and TRAE all broke out there. The reasons are straightforward in the article: code is highly standardized, the digital loop is closed, and error feedback is clear.
But coding is only one step in a workflow. The broader market is office work: reading email, attending meetings, searching documents, handling data, processing approvals and following up on decisions. The article contrasts a market of tens of millions of developers with a general office market of billions of workers, and says the latter has a token consumption ceiling that is orders of magnitude higher.
The same pattern can be seen globally. After custom agents cooled off, OpenAI shifted focus to ChatGPT Projects and Operator, pushing scattered functions back into one unified window. Microsoft reworked Copilot from a standalone Office add-on into a single interface that runs across Microsoft 365 workflows. Salesforce’s Agentforce has also moved away from emphasizing the flexibility of a single agent and toward unified scheduling inside enterprise CRM systems.
The domestic market is accelerating as well. The article says Tencent’s WorkBuddy, a strategic product that Tencent has high expectations for, has seen monthly active users and daily active users climb quickly in a short period and has already become the most active AI agent application in China’s efficiency category. It also notes that Kingsoft Office recently released products with a similar positioning: Lingxi Professional Edition for individuals and WPS Comate for organizations, both aimed at bringing AI agents deeper into documents and office scenarios.
The article’s conclusion is that a broad consensus is taking shape globally: scattered agent entry points are fading, and the super workbench is replacing them.
What is being rewritten is work itself
The article cautions against reading this competition as a simple replay of the Office software era. For the past two decades, enterprise software architecture has largely been split across ERP, CRM, OA, HR, finance and project management systems, each guarding its own domain. Employees moved from one system to another to get things done.
In that setup, people were the ones stitching everything together. In the future, the article argues, agents will take on that role.
That is also why Tencent, Alibaba and ByteDance have all handed this round of consolidation to cloud and collaboration teams. An agent’s execution efficiency, the piece says, depends less on how smart the model is and more on how much enterprise data and how many API interfaces it can orchestrate.
Alibaba’s planned “Qwen Office” would sit on top of DingTalk’s enterprise relationship graph, organizational structure and approval flows. Tencent’s WorkBuddy sits on the collaboration ecosystem of WeChat and Tencent Docs. ByteDance’s TRAE Work is backed by Feishu’s knowledge base and workflow engine.
In that view, consolidating agents is really a battle for absolute control over enterprise data and system API orchestration. Whoever becomes the first AI interface employees open each day could end up controlling the dispatch center for enterprise data and capability. Once users only need a single super workbench entry point, thousands of other software products lose their reason to be opened directly.
Software moves into the background
One of the arguments repeated across the industry over the past year is that AI is killing SaaS. The article pushes back on a full rewrite scenario. No company, it says, is ready to hand over back-end logic built over more than a decade, logic tied to compliance and core assets, to a large model for complete reconstruction.
In the past, the core premium of SaaS came from the interface and the workflow built around it. Vendors designed buttons, menus and forms carefully, and employees adapted to the software’s logic.
When agents unify the workplace entry point, that rule changes. The piece says the key driver is the spread of Skills, described as skill or capability interfaces.
Software no longer has to show humans a complicated UI. It only needs to connect its own Skills to the super workbench. That could mean a supply chain inventory query in SAP, a customer profile pull inside Salesforce, or the generation of an expense reimbursement voucher in Kingdee. Functions once buried under layers of submenus are packaged into standardized Skills.
The article says this shift will rebuild the layering of enterprise services. The front end belongs to the super workbench. The back end belongs to software. The room left for many vertical agents in the middle gets smaller. If employees no longer open SaaS interfaces directly, the interface premium behind per-seat pricing also weakens. Software vendors would then move from selling UI to charging based on how often Skills are called and what results they deliver.
In the article’s words, the greatest value of enterprise software over the past two decades was the interface; the greatest value of agents is making the interface disappear. Skills are the bridge between the two.
The best agent may be invisible
The article closes by dividing agent evolution into three stages.
- Stage one: the agent is a product, with many independent forms competing at once.
- Stage two: the agent becomes an entry point, and large companies focus resources on winning the main interface for work and the operating system.
- Stage three: the agent becomes a capability, present everywhere but no longer visible as a distinct product.
In the authors’ view, the latest consolidation by Tencent, Alibaba and ByteDance marks a move from stage one into stage two, while the industry is also sliding quickly toward stage three. Agents are going through another round of capability sink, much like browsers and super apps did earlier in internet history. They move from being highly visible star products to underlying infrastructure, like electricity or the network itself.
The article says agents may not become the next WeChat, but they could become the next Windows.
In an afterword section, the authors extend that point: technology often begins as something new, then becomes a product, then becomes a capability, and eventually becomes “nothing” in the sense that people stop naming it as a standalone thing. Electricity has no separate entry point. Network protocols have no separate entry point. Databases have no separate entry point. Their disappearance does not mean failure; it means they have become foundational. By that logic, the authors suggest people may stop talking about agents a few years from now, much as few people talk explicitly about HTTP today. A mature foundational technology, they write, eventually loses its name.
This article was originally published on the WeChat public account “版面之外” and credited to the authors Huahua and Banjun.