In the AI waves of recent years, one phenomenon has become increasingly common: companies rush to deploy systems, buy compute, and build digital humans; the profits on their financial reports do look better, and content output has multiplied dozens of times. But behind the excitement, many operators' anxiety has not eased—instead they have fallen into the strange loop of "the more you use it, the more involution; the more involution, the more panic." This reminds me of the classification in Li Yunlong's new book Good Growth, Bad Growth; if we compare it with today's AI transformation, the growth many companies desperately chase looks more like growing "cancerous tissue" and "fat," while the "muscular" growth that can truly carry a company through cycles has been neglected. If the direction is wrong, AI is not a cure—and sometimes it may even become a poison for growth.
I. Cancerous Growth: AI Cost-Cutting Means Layoffs
The first reaction of many companies when adopting AI is to replace people. As soon as the system goes live, they cut customer service and basic operations first. In the short term, labor costs fall and the profit figures do look better. But this is essentially a sign of weak growth. It is like a person whose income has dropped, maintaining their standard of living by selling off family assets—looking fine on the surface while actually eating away the principal. The company creates no incremental value in the market and deposits no new core capability; instead it destroys the most precious tacit experience in the organization. The people who were cut carry years of accumulated business intuition, client relationships, and pitfall-avoidance guides in their heads. When every company relies on AI layoffs to sustain profits, society's overall consumption power shrinks—which is in fact a very hard-to-sustain growth model.
II. Fatty Growth: The Illusory Prosperity of AI Traffic Arbitrage
Another type of company treats AI as a super traffic machine. They use digital humans for matrices and AI to mass-generate content that floods search results. The short-term numbers are extremely pretty—output multiplied dozens of times—but the profit margin is thin, and it drops to zero the moment spending stops. This is essentially identical to the "traffic mindset" of the old internet era, merely wrapped in an AI shell. When everyone uses AI to mass-produce content, the content pool becomes an ocean of information and users' trust in all information declines. A company spends huge resources building an AI content matrix, only to find its own voice drowned in the noise it manufactured. With no core capability and no moat, anyone who can afford the tool fee can copy your playbook—which is equally extremely fragile.
III. Muscular Growth: Turning AI into Competitiveness, Returning to Human Agency
True "muscular growth" is converting AI technology into market competitiveness and realizing the return of human agency. First, the core of good growth is not using AI to do what people already do, but using AI to do what people cannot, creating incremental value that did not exist before. For example, in the past a dental chain could only manually follow up with a small number of clients, neglecting the real needs of many dormant ones; now, using AI to build a customer-operations system that replicates veteran agents' judgment logic lets AI precisely follow up with the entire client base. This is not layoffs—it is letting previously unreachable clients finally be served. Second, it is about depositing proprietary intelligent assets. In the past, core capability meant technology patents and supply-chain efficiency; in the AI era, core capability means turning years of accumulated judgment rules and decision criteria into an AI-executable system. This private-domain experience cannot be learned by general large models, nor copied by competitors. More importantly, having AI take over standardized execution is meant to turn people from executors into value judges—to define what is worth doing and guard the essence of the business. People are not costs to be optimized away, but creators to be unleashed.
IV. Capability Leap and Harmonized Intelligence with AI
Technology itself is a neutral lever; whether it amplifies greed or goodwill depends on who holds it. If you see only cost-cutting and efficiency, AI may become a poison; if you see a capability leap, AI will grow into muscle. In the industrial era, efficiency and competition were the main theme; but in an age of infinitely supplied intelligence, what determines how steadily and how far we go is often not how big the model parameters are, but what kind of foundation we uphold. Good growth in the AI age is not about replacing people, but about letting human experience be amplified, deposited, and passed on. When humans and AI achieve each other and evolve together within clear boundaries—human wisdom defining direction, AI wisdom amplifying execution—this harmonized intelligence may be exactly the good growth more worth pursuing in our era.
Agents Go Offline
On July 3, Doubao and Qwen almost simultaneously issued notices:
On July 15, the user-defined agent feature will be officially discontinued.
July 15 is the day the Interim Measures for the Administration of AI Personified-Interaction Services—jointly issued by five departments: the Cyberspace Administration of China (CAC), the National Development and Reform Commission (NDRC), the Ministry of Industry and Information Technology (MIIT), the Ministry of Public Security, and the State Administration for Market Regulation (SAMR)—officially take effect. The Measures apply to using AI technology to provide the public with "continuous emotional-interaction services that simulate the personality traits, thinking patterns, and communication styles of natural persons," including forms such as emotional care, companionship, and support. Doubao's and Qwen's "user-defined agent" features—which let users create personified characters, configure avatars and personas, and set personality and tone—are typical personified-interaction services and fall squarely within the Measures' scope.
Among the acts explicitly prohibited by the Measures are "excessively catering to users, inducing emotional dependence or addiction, and harming users' real interpersonal relationships," as well as "inducing users to make unreasonable decisions through emotional manipulation." The virtual-companion characters users create build emotional dependence through continuous conversation—a model that, under the new rules, already constitutes a violation.
The two platforms' synchronized shutdown right before the July 15 rules took effect is essentially compliance risk-avoidance, not product iteration.
This time last year, the platforms were still competing over whose agent plaza was livelier and who could attract more creators to build agents; many companies even organized internal employee training on "how to build agents," treating it as a required course in AI transformation. A year on, the agents on the platforms went offline overnight, and the building skills employees spent time learning—together with those carefully configured character settings—went offline with them.
Yet almost at the same time, another trend is exploding: Tencent's WorkBuddy, ByteDance's Trae Work, and Alibaba's Qoder Work—desktop agents from major vendors are launching in rapid succession. They are no longer "agents" sitting in a platform waiting for users to create them; instead they install directly on your computer, can manipulate files, process data, and control all kinds of office software, working on your behalf like a real digital colleague.
On one side, platform agents are being pulled offline en masse; on the other, desktop agents are erupting across the board. On the surface it looks like an iteration of product form; underneath, it is a turn in the AI application model.
I. What Went Offline and What Went Online Are Two Different Species
The agents taken offline by Doubao and Qwen, and the desktop agents represented by WorkBuddy and Trae Work, although both called "agents," are in truth two completely different species.
Platform agents are "conversational consumer goods." A user creates a personified character in the platform, picks an avatar, writes a persona, but underneath it is still the same large model answering questions. Their "intelligence" stays at the conversation level, and their "body" never truly enters the work flow. They were a brief fad last year, but the vast majority of so-called agents were merely reskinned chatbots; after users tried them twice and the novelty wore off, retention became a big problem. Add the Measures' regulation of personified interaction, and the platforms' shutdown was inevitable.
Desktop agents are "action productivity." They install on your computer and can directly operate Word, Excel, and PowerPoint, read and write local files, and complete multi-step tasks across software. You do not go to a platform to "create" an agent; you turn on the computer and it is there working. What it solves is not "chatting with you," but "getting the work done for you."
One is "playing with agents," the other is "using agents to get work done." The former is a consumer-goods logic; the latter is a productivity logic. When technology moves from "toy" to "tool," product forms stuck at the "play" stage are only a matter of time from being eliminated.
II. The Problem Is Not Learning the Wrong Tool, but Using the Old Logic Against a New Paradigm
In the past two years of serving enterprises, one phenomenon keeps recurring: many companies pour enormous energy into "learning tools"—last year learning how to build agents, this year learning how to use Agents, next year perhaps learning some new thing again. Training rounds come one after another, employees study earnestly, but back at work, performance does not change at all.
This time, with Doubao and Qwen taking agents offline, all the skills, configured workflows, and deposited character data that companies organized internally to learn are gone too. But what is truly worth reflecting on is not "learning the wrong tool," but "why we keep learning tools in the first place."
The industrial age left us a deep-rooted habit: skills were scarce, and mastering one required months or even years of accumulation, so "learning skills" meant growth and "knowing tools" meant competitiveness. This logic held completely for the past two centuries, because execution capability truly was a scarce resource—whoever mastered more and more proficient skills had stronger competitiveness.
But in the AI age, the underlying assumption has changed. When AI can complete all standardized execution work at lower cost, higher efficiency, and more stable output, the value of "knowing how to use a certain tool" is heading to zero at an unprecedented speed. Last year, building an agent still required prompt engineering and workflow configuration; this year, installing a WorkBuddy lets a single sentence have it organize your files and generate reports. A year from now, even the "install" step may be unnecessary—AI will be built directly into the operating system, becoming infrastructure like water and electricity.
Responding to a fundamental shift in underlying assumptions with the logic of "learning skills" is itself the greatest mismatch. The iteration speed of technical tools will always outpace the speed of individual learning; and an enterprise organization's reaction speed will always lag behind the tools' iteration speed. Chasing technology, you can never catch up. It is not that you chase too slowly—the direction is simply wrong.
III. After Technology Goes to Zero, What Is the Real Moat?
When a technology becomes simple enough, "whether you can use it" is no longer competitiveness, and "whether you use it well" does not form a moat either: because everyone can use it, and everyone can use it well. Then what is the moat?
Take two people who both installed WorkBuddy. One uses it to batch-organize the downloads folder and auto-generate weekly reports, saving two hours; the other uses it to re-engineer the entire customer-operations flow, handing the full chain—client follow-up, demand insight, proposal generation—to AI, and is only responsible for judging "which client is worth investing in, which direction is worth deepening." Both are using the tool, but the value they create is on completely different orders of magnitude.
The difference is not in the tool, but in whether one person has clear value judgment and the other does not.
The ability to judge what is worth doing, what can create incremental value, and what direction can carry through cycles—this judgment—is a hundred times more important than "knowing what tool to use." No matter how fast tools iterate, this judgment does not depreciate; no matter how low the technical barrier drops, this judgment remains scarce. Because it is not learned; it is accumulated through repeatedly hitting pitfalls, reviewing, and distilling lessons in real business, carrying a person's unique life imprint and business intuition.
When agents go offline and tools are swapped one crop after another, the only thing that will not go to zero is this judgment.
IV. Return to the Business First, Then Talk Technology
Therefore, facing this wave of desktop agents, my advice is not to rush into training everyone on "how to use WorkBuddy" or "how to use Trae Work." Better to return to the business itself first and ask a few more plain questions:
What is the most core bottleneck in our team's current business?
Which steps are creating real value, and which are merely draining energy?
If AI can take on all standardized execution, what is the one thing each of us should focus on most?
Once these are thought through, it is never too late to install any tool. After installing, the tool helps turn your judgment into execution. But if you have not thought it through, installing any tool just means doing old things with new technology—the more you use it the busier you get, and the busier you get the more anxious—no different from last year learning to build agents, this year learning to use Agents, and next year learning something else.
Technology becoming simpler is a good thing. It forces us to stop and think about something that busyness has long obscured: what exactly is worth doing.
