Over the past month, Tencent, Alibaba, and ByteDance have suddenly begun doing the same thing.
Instead of launching new Agents, they've started scaling back Agents.
Tencent issued an internal notice. All business units and some teams from Tencent's QClaw Product Center are being merged into WorkBuddy. WorkBuddy is touted as having the potential to become the third blockbuster product after QQ and WeChat. Just consider the weight of that evaluation.
Alibaba has also moved, with even bigger actions. According to information obtained by Caijing, Alibaba is about to launch "Qianwen Office", integrating three agent products—QoderWork, Wukong, and MuleRun—into it at once. The person in charge is Chen Yusen, the new CEO of DingTalk.
Just handed to him in early July this year, these three products have now been consolidated into one. 'Qianwen' is currently Alibaba's most well-known AI brand. Stuffing the office products uniformly into this system makes their intention crystal clear.
ByteDance is also busy. Its AI programming product TRAE SOLO quietly changed its name: TRAE Work. A change of one word shifts the direction by 180 degrees.
Looking at each case individually, they are just ordinary organizational adjustments.
But with three giants acting almost simultaneously, converging the slew of Agents they heavily promoted over the past half-year into a single entry point, this is no coincidence.
This is the first genuine product consensus of the Agent era.
I. We Don't Need So Many Agents Anymore
Think back to the initial explosion of Agents, what was the spectacle like?
Around Chinese New Year this year. Almost every company thought Agents would be like the WeChat Official Accounts back in the day.
One team makes one, one department makes one, one scenario makes one. Everyone frantically places bets, no one knows who will ultimately win.
In just a few months, thousands of agents with similar functionalities emerged.
Technological barriers instantly dropped to zero.
Tencent is a typical 'shrimp farmer'. QClaw was built by the PC Manager team based on OpenClaw, WorkBuddy was developed by Tencent Cloud, and there are also QQ Lobster, Browser Lobster... different versions of 'Lobsters' scattered across various business groups, each fighting its own battles.
Multiple Agent projects were running simultaneously within Alibaba. The desktop-level AI agent tool QoderWork remained in a semi-finished state; Wukong, which had long been in refinement; and MuleRun, mainly targeting overseas markets. Product positioning was similar, functionalities heavily overlapped, computing power and R&D resources were fragmented, users were left confused, failing to form a unified perception.
ByteDance didn't fall behind either, launching several 'Lobster' Agents in succession—ArkClaw, ByteClaw, Feishu aily, Trae SOLO—spread out like deploying troops.
Frankly speaking, this dense betting isn't a characteristic unique to Chinese companies.
OpenAI, Anthropic, Google, all followed the same playbook. Operator, Deep Research, Canvas, Projects, Codex, Claude Code, NotebookLM, Gemini... Each company's product lines popped up like chimneys.
This isn't a sudden phenomenon. When something new emerges, during the exploration phase, the best management is to allow duplication.
But trial and error costs money.
Over the past six months, major tech companies running multiple Agent lines have consumed vast amounts of inference computing power. Multi-step reasoning, repeated API calls, redundant context retrieval made the single-run cost of an Agent far exceed that of ordinary conversational models. High costs persisted, while enterprise-side payments didn't keep up. Dispersed product lines dragged down the overall return on investment.
After open-source tools leveled the technological playing field, the core of competition changed. Computing power budgets couldn't sustain unlimited internal consumption. Resources had to be concentrated.
When the direction became clear, exploration ended.
II. Big Tech Starts Actively Killing Their Own Innovation
That sounds contradictory.
Innovation is a good thing, why cut it off with your own hands?
But if you look at the longer history of the internet, this has happened repeatedly. The entire internet has essentially done three major things to unify entry points.
PC Era: Browsers unified all web pages. Mobile Era: Super Apps unified all services. AI Era: Super Workbenches are unifying all Agents.
How many social products has Tencent made?
WeChat, QQ, Qzone, Pengyou.com, Tencent Weibo... In the end? One WeChat consolidated the entire game, becoming the super entry point of the mobile internet. Meituan fought in group buying, movies, food delivery, hotel & travel, ride-hailing, and eventually became one Super App. Didi operated express rides, premium rides, carpooling, designated driving, and eventually packed them all into one Didi.
Every time products converge, it's not because innovation failed, but because the market gradually moves toward certainty. The true sign of a market maturing is never more and more products, but fewer and fewer.
All platform wars, in the end, are not fought with the ability to create, but with the ability to subtract.
Look at how these three giants are clearing the battlefield.
Tencent is taking the exploratory results scattered within the PC Manager team, returning them to CSIG (which focuses on cloud services), packing them into WorkBuddy, and placing command under Tencent Cloud.
Alibaba is formally abandoning the decentralized approach of free incubation of AI assistants by various business lines. Command over office AI is being comprehensively handed over to 'Qianwen Office' led by DingTalk, with the brand uniformly incorporated into the 'Qianwen' system.
ByteDance is shifting TRAE towards workflow collaboration, effectively killing the era of Agents as independent soldiers. SOLO has become the invisible technical foundation within TRAE Work.
In internet history, every product unification signifies the real war is beginning.
III. Programmers Are No Longer the Biggest Market
Beneath this consensus to consolidate lies an even deeper shift.
Why did TRAE SOLO have to change its name to TRAE Work? Why does Claude look less and less like a chatbot? Why is OpenAI frantically integrating Chat, Code, Research, Projects, Operator? What's the goal?
The answer is straightforward.
For the past year, everyone thought the biggest market for AI was programmers.
Now, big tech is starting to realize, it's not.
Programmers are just AI's first batch of users. The truly massive market is all working people.
Undeniably, the Coding scenario was the first to be validated by the industry. Cursor, Claude Code, TRAE all broke through. The reason isn't hard to understand: code is highly standardized, forms a digital closed loop, and has clear error tolerance and feedback.
But writing code is just one link in the workflow. The scenario that truly covers everyone is reading emails, attending meetings, checking documents, processing data, approvals, and following up on decisions.
The market for writing code is tens of millions of developers. The market for general office work is billions of workplace individuals. The latter's volume and token consumption ceiling are hundreds and thousands of times higher than the former.
Not only that, today, global giants are almost all converging in the same direction.
After custom agents met with lukewarm reception, OpenAI quickly shifted focus to ChatGPT Projects and Operator, compressing scattered features back into a unified window.
Microsoft restructured the Copilot architecture, upgrading it from a lonely Office add-in to a super entry point running through the entire Microsoft 365 workflow.
Salesforce's Agentforce no longer emphasizes how flexible a single agent is, but instead emphasizes unified scheduling capabilities within the enterprise CRM.
The domestic market is also accelerating. Tencent's WorkBuddy, as a strategic-level product Tencent pins high hopes on, has seen its Monthly Active Users (MAU) and Daily Active Users rapidly climb in a short time, becoming the application with the highest activity among domestic efficiency-oriented AI agents.
Even traditional enterprise service and office giants can't sit still. Kingsoft Office (WPS) also recently launched similarly positioned products: Lingxi Pro for individuals and WPS Comate for organizations, attempting to fully integrate AI agents into document and office scenarios.
Thus, the gears of business logic have turned.
A global consensus is forming: as dispersed Agent independent entry points gradually disappear, they are being replaced by the Super Workbench.
IV. What's Truly Being Rewritten is Work
These three giants are clearly vying for the territory of the Super Workbench, but don't mistake this battle for a rehash of traditional Office.
How to understand this?
Over the past two decades, the operational model of enterprise software architecture has been: ERP, CRM, OA, HR, Finance, Project Management—each system guarding its own plot. Employees jump back and forth between these systems.
What strings everything together at this stage is the human. In the future, what strings these systems together will be Agents.
This is also why Tencent, Alibaba, and ByteDance have, without prior agreement, handed this round of consolidation to their cloud and collaboration office teams.
An Agent's execution efficiency doesn't depend on how smart the model itself is, but on how much enterprise underlying data and API interfaces it can orchestrate.
Alibaba's upcoming 'Qianwen Office' relies on DingTalk's enterprise relationship chain, organizational structure, and approval workflows. Tencent's WorkBuddy relies on the collaboration ecosystem of WeChat and Tencent Docs. ByteDance's TRAE Work relies on Feishu's knowledge base and workflow engine.
The giants' consolidation of Agents is, at its core, a fight for absolute orchestration rights over these enterprise data and system APIs. Whoever becomes the first AI entry point an employee opens every day will hold the control center for orchestrating all enterprise data and capabilities.
When users only need one Super Workbench entry point, the remaining thousands of software lose their reason to be opened directly.
V. Software Begins Receding to the Background
Over the past year, there's been a hotly debated view in the industry: AI is killing SaaS.
People thought Agents would sweep in like a storm, directly overturning and rewriting the ERP, CRM, HRM, and financial software that enterprises have used for over a decade.
The reality is, no enterprise dares to hand over the backend logic—accumulated over decades, tied to compliance and core assets—entirely to large language models for reconstruction.
In the past, SaaS's core premium came from interactive interfaces and user workflows. Vendors meticulously designed every button, menu, and form, making employees adapt to the software's logic.
But when Agents unify the work entry point, these rules are overturned. What truly changes this is the proliferation of Skills (capability interfaces).
Software no longer needs to present complex UIs to humans; it only needs to connect its own Skills to the Super Workbench—whether it's querying supply chain inventory in SAP, retrieving customer profiles in Salesforce, or generating financial reimbursement vouchers in Kingdee.
These capabilities originally hidden deep within dozens of sub-menus are packaged into standardized Skills.
This transformation will directly reshape the industry layers of enterprise services. The frontend belongs to the Super Workbench, the backend belongs to the software. The space for numerous vertical Agents in the middle will shrink.
When employees no longer open the SaaS interface, the per-seat interface premium will also become ineffective. Software vendors will be forced to shift from selling UI interfaces to charging based on the frequency of Skills calls and the outcomes delivered.
The greatest value of enterprise software over the past two decades was the interface. The greatest value of Agents is making the interface disappear.
And the bridge connecting the two is precisely Skills.
VI. The Best Agent is an Invisible Agent
Looking back at the evolution trajectory of Agents, three stages are clear.
Stage One: Agent as a Product. Everyone was making them, forms were diverse, chimneys were erected densely.
Stage Two: Agent as an Entry Point. Big tech begins consolidating resources, fighting fiercely for the main entry point of work and operating systems.
Stage Three: Agent as a Capability. Entry point restructuring is complete, it's omnipresent, yet formless and shapeless.
Today's consolidation actions by Tencent, Alibaba, and ByteDance mark the industry formally moving from Stage One into Stage Two, and rapidly sliding toward Stage Three.
Agents are experiencing another capability descent in internet history, following browsers and Super Apps. They are transforming from dazzling star products into essential yet unmentioned underlying infrastructure, like electricity and the internet.
Agents won't ultimately become the new WeChat, but they are highly likely to become the new Windows.
Words from [Beyond the Layout]:
The development of technology always goes through an interesting process.
At first, it's a new thing. Later, it becomes a product. Then, it becomes a capability. Ultimately, it becomes nothing.
Electricity has no entry point of its own, network protocols have no entry point of their own, databases have no entry point of their own.
Their disappearance doesn't signify failure. It's because they've become the foundation of the entire world.
Perhaps in a few years, we won't discuss Agents anymore either. Just like today, no one discusses HTTP.
Truly mature underlying technologies eventually lose their own names.
This article is from the WeChat public account "Beyond the Layout", authors: Hua Hua, Ban Jun
