An edited English Q&A with Shenzhen TV on what a viral Midea PortaSplit story reveals about AI agents, loop engineering, agent societies, and the human ability to organize digital labor.
This post is adapted from a Shenzhen TV interview with Benyou Wang on AI agents. The discussion began with a surprisingly vivid case: Chinese air conditioners, especially Midea’s PortaSplit, became highly sought after in Europe during a heat wave. According to public reports, Austrian resident Denis Yurchak spent two days searching for a unit, monitored stock with AI agents, and eventually drove about 200 kilometers to secure one. The case was covered by Xinhua, and Yurchak also shared the experience on X.
At first glance, this looks like a consumer story: a shortage, a popular product, and someone determined enough to chase inventory across platforms and cities. But it also marks something more interesting. Many earlier AI agent examples were framed around coding, slide generation, report writing, or office automation. This case spread because an agent helped an ordinary person solve an ordinary but urgent life problem: finding a real product, in a real market, under real constraints.
The deeper lesson is not that AI replaced the human. The lesson is that the human who succeeded was the one who knew how to organize AI.
Shenzhen TV: Why did this air conditioner case attract so much attention in the AI community?
Benyou Wang: Because it moved AI agents out of the demo room.
For a long time, many agent examples were about knowledge work: writing code, preparing slides, generating reports, or automating office routines. Those are important, but they still feel close to the traditional world of software demos. This case was different. A person had a concrete need. The product was out of stock. The information was scattered across retailers, countries, languages, and logistics options. The user had to monitor inventory, compare alternatives, communicate across channels, and make decisions quickly.
That is closer to project management than to a simple app operation.
Ordering a drink, booking a hotel, or buying an airline ticket is usually a standardized workflow. The interface is known, the options are structured, and payment is the final step. In the PortaSplit case, the task was open-ended: search multiple sources, watch for inventory changes, make contingency plans, and decide when human action should take over.
This is why the story is valuable. It shows that AI agents are becoming useful not only when they generate content, but when they coordinate action under uncertainty.
Shenzhen TV: AI agents are becoming better at completing tasks. But in this case, the real winner still seems to be the person who knew how to organize AI. In the future, will competition between people increasingly become a competition in the ability to organize AI, rather than only knowledge and skills?
Benyou Wang: Yes, but we should define “organizing AI” carefully.
It is not just prompting. It is the ability to choose the right task, place the agent in the right environment, provide the right tools, create feedback, and know when to intervene. I often call this loop engineering: building a loop in which the system can observe the current state, act through tools, verify the result, revise the plan, and repeat until a clear completion condition is reached.
AI will not automatically make a lazy person strong. It rewards people who can decompose a problem, design a process, inspect outputs, and keep revising. A good user does not simply say, “Do this for me.” A good user creates a working relationship with AI: define the goal, let the agent act, check the result, correct the direction, and improve the next iteration.
This is also why expert involvement remains important. In many real tasks, the value comes from the interaction between AI and experts. AI can search, draft, compute, and monitor. Experts know what matters, what is risky, and what should count as a good answer. The best outcomes often emerge from an AI-expert loop rather than from either side alone.
There is also a human leadership problem. Our cognitive bandwidth is limited. Opening ten AI chat windows does not automatically produce ten times the intelligence. Someone still has to decide priorities, allocate subtasks, compare results, resolve conflicts, and take responsibility for the final decision. In that sense, the future may require a new kind of leadership: the ability to organize digital labor while still exercising human judgment.
Shenzhen TV: If everyone has their own AI agents, platforms will no longer face only human users. They will face thousands or millions of agents acting on behalf of users. Do existing internet rules and business models need to be redesigned around an “agent society”?
Benyou Wang: I think many rules will have to be rewritten.
The first issue is the chain of responsibility. When an agent searches, negotiates, reserves, purchases, or communicates on behalf of a user, we need to ask: who authorized the action, who verified it, who is liable for mistakes, and what evidence should be preserved? Today’s internet was designed mainly around human clicks. Agent actions require more explicit permission, logging, auditability, and fallback.
The second change is the platform entrance. In the past, companies competed for traffic entrances: search engines, app stores, social feeds, and recommendation systems. In an agent society, platforms may compete for proxy entrances. If an AI assistant becomes the layer that chooses which restaurant, product, hotel, or service to recommend, then businesses will increasingly care about how they are represented to models and agents.
This will change advertising and search. Traditional ads may be ignored by agents if they do not help the user’s objective. We may see more attention to GEO, or generative engine optimization: making products, services, and facts understandable, trustworthy, and retrievable by AI systems. Some advertising may even target the model-mediated decision layer rather than the human eyeball directly.
The third change is resource coordination. If every user can deploy many agents, platforms will face new questions around rate limits, identity, fairness, and anti-abuse. Otherwise we may get an “infinite monkeys” problem: many agents repeatedly probing systems, consuming tokens, scraping inventories, or creating artificial demand. Agent societies need rules for identity, access, quotas, and coordination.
Finally, we should not imagine this as a machine-only society. Human-in-the-loop mechanisms will remain important. For high-stakes decisions, the agent should not simply act. It should surface evidence, ask for authorization, and know when to hand control back to people.
Shenzhen TV: Some people believe AI will give ordinary people expert-like ability. Others argue that people with stronger learning and technical ability will benefit more. How do you see this?
Benyou Wang: AI is both an equalizer and an amplifier.
It lowers the barrier. More people can write code, design products, analyze data, prepare documents, search literature, and build prototypes. In this sense, AI raises the floor. Many tasks that once required specialized training become accessible to more people.
But AI also raises the ceiling. People who already know how to learn, ask good questions, verify answers, and organize workflows can use AI more effectively. They can turn the same model into a research assistant, a coding partner, a product analyst, a data engineer, or a project coordinator. The tool is available to everyone, but the ability to extract value from it is not evenly distributed.
In professional domains, professionals still have advantages. AI can draft a medical explanation, legal memo, research summary, or technical plan, but experts know the hidden assumptions, failure modes, and boundaries of acceptable reasoning. The stronger the domain, the more important expert judgment becomes.
So AI will democratize access to capability, but it will also magnify differences in taste, judgment, discipline, and learning speed. It is a capability amplifier. It helps more people reach a useful baseline, while helping the best people move even faster.
Shenzhen TV: In the past, experience came from long accumulation. Today AI can quickly integrate information and generate plans. Will the advantage of human experience be redefined? Which abilities will become more irreplaceable?
Benyou Wang: Experience will not disappear, but its meaning will change.
In the past, experience often meant remembering the road. In the future, experience may mean knowing what to do when the map fails.
AI can retrieve information quickly, compare options, and produce a plan. But it may still misunderstand the real goal, miss a social constraint, trust a noisy signal, or fail to notice that the environment has changed. Human experience becomes valuable when the situation is ambiguous, when the feedback is delayed, or when the cost of a mistake is high.
Several abilities will become more important.
The first is product sense: knowing what AI should be used for. Many projects fail not because AI cannot do anything, but because no one has defined a meaningful task. The key question is not “What can this model do?” but “Where is the most expensive, frequent, measurable friction in this workflow?”
The second is technical judgment. Even when AI can generate a solution, someone must judge whether the solution is robust, maintainable, and appropriate for the environment.
The third is leadership, empathy, and responsibility. Real AI systems touch people. They change work, expectations, trust, and risk. A good leader must organize resources, communicate across disciplines, understand users, and remain accountable for outcomes.
The fourth is critical thinking. If humans stop asking better questions, AI has little hope of producing better answers. Critical thinking does not become less important in the AI era; it becomes the steering wheel.
For universities, this has educational significance. We should not only train students to use tools. We should educate people who can think critically, communicate across languages and cultures, take responsibility, and organize knowledge for real public value. In an AI era, whole-person education becomes more important, not less.
Shenzhen TV: Many models now claim to have agent abilities. Some agents can order milk tea, book meals, buy flight tickets, or reserve hotels. But these are standardized tasks. The Austrian PortaSplit case looks more like project management: multi-platform search, inventory monitoring, cross-border logistics, multilingual communication, contingency planning, and human-AI decision making. What core abilities should a mature agent have? What are the industry’s shortfalls?
Benyou Wang: The key shortfall is fallback ability.
Current agents show what I call jagged intelligence. Their reasoning, planning, and tool use do not mature along one smooth curve. They are strong in some local regions and fragile in others. A model may perform impressively on a task that looks difficult, then make a basic mistake on a nearby task that looks easier. The boundary of capability is uneven and hard to predict from human intuition.
That means the delivered product cannot simply expose the raw model’s jaggedness. A mature agent system must wrap the model with tools, context, validators, permissions, retry logic, and human takeover mechanisms. This broader system is what I call harness engineering. The model may have uneven capability, but the product must provide more stable reliability.
A mature agent should have at least six capabilities.
First, it should observe state. It needs to know what has already happened, what information has changed, and what remains uncertain.
Second, it should use tools reliably. Searching, checking inventory, sending messages, editing files, querying databases, and making reservations must happen through controlled interfaces.
Third, it should verify outcomes. It should not assume that an action succeeded just because it called a tool. It must check confirmation signals, logs, receipts, database states, or other evidence.
Fourth, it should replan. When one path fails, it should not collapse. It should compare alternatives and update the plan.
Fifth, it should contain errors. Some actions should be reversible, some should require confirmation, and some should trigger a human handoff.
Sixth, it should know when to stop. An agent without a stopping rule can become dangerous, expensive, or simply useless.
This is why closed loops matter. In multi-step tasks, reliability is multiplicative rather than additive. If each step is 99 percent correct, a 50-step workflow has only about a 60 percent chance of being entirely correct without correction. Strong single-step capability does not guarantee end-to-end reliability.
The solution is not only a bigger model. The solution is loop engineering: the system acts, checks, receives feedback, repairs, and tries again. Software engineering became an early mature agent scenario partly because code offers natural feedback. The object is digital, tools can execute actions, and compilers or tests can detect many errors.
Many real-world domains are harder. In healthcare, manufacturing, public services, urban governance, embodied intelligence, and AI for science, feedback may be delayed, partial, expensive, or irreversible. Before agents can scale there, we must turn open, messy processes into local closed loops that can be observed, verified, and controlled.
This is why I am optimistic about AGI, but in a specific sense. AGI will not arrive uniformly everywhere at once. It will first appear as local AGI in bounded environments where data flows, tools, and verification mechanisms are strong. Coding, mathematics, data analysis, research support, and operations may move quickly. AI for science, embodied systems, and medicine may take longer because their loop engineering and environmental harnesses are not yet complete.
Shenzhen TV: For traditional enterprises considering AI agents, how can they tell whether they truly need AI or are simply being swept up by AI anxiety?
Benyou Wang: I suggest a simple counterfactual test.
Ask: if AI did not exist today, would this business problem still be worth solving?
If the answer is no, then the project probably comes from AI anxiety rather than business need. If the answer is yes, AI may be a useful tool, but the company should still start from the problem rather than the model.
Good AI scenarios are usually painful before AI arrives. They involve repeated work, high labor cost, slow response, frequent errors, scattered knowledge, or too much time spent searching, checking, and moving information between systems.
Then we should ask whether the workflow has the right conditions:
If a process has no standard, the data is poorly governed, and the systems are not connected, the first step is often not deploying an agent. The first step is process mapping, data governance, and interface construction.
AI cannot automatically repair an undefined organizational process. It will amplify both the strengths and the confusion already present in the organization.
This is why I believe 2026 is a turning point for AI agents in China. The field is moving from model capability competition toward closed-loop engineering and scalable delivery. The most important question is no longer just “How strong is the model?” It is:
Can we build an environment where the agent can act, be checked, correct itself, and hand control back to humans when necessary?
That is the practical path from impressive demos to dependable AI systems.