Let AI do things for you, but I hope you can still keep learning

Over the past half month, I've written quite a few articles and intensively shared a lot of my own practice. This content involves many AI tools; the domain span looks large, but underneath it all, efficient collaboration is achieved through Agent collaboration and Skill engineering.

These days, apart from visual and interaction prototype refinement, which for now still can't do without Figma (many workflows have also been deeply streamlined and automated by AI), almost all stages—research, writing, programming, and so on—can be assisted and completed by AI. After six months of high-intensity tinkering with AI, I think it's time to take stock: stepping away from that omnipotent input box and truly looking inward, what exactly have I learned, and how has my understanding evolved.

Agents have truly become very powerful this year

Looking back at the second half of last year, Agents had just become popular globally. At that time, most forms stayed at the level of Dify or building on frameworks like LangChain—essentially still "prompts + tool calls + model APIs." Complex tasks relied on manually orchestrated Workflows and lacked genuine autonomous planning capability.

At that time I was using a coding Agent from ByteDance—Trae, because I really like ByteDance, so I chose it among Code Buddy and Qoder. Agents back then were generally clumsy: no binding specification documents, and certainly no mature Skill ecosystem; planning relied entirely on the model's built-in Todo mechanism. Most of the code written could only be called "barely running"; shipping to production was completely out of the question.

But this year, with the maturation of the Skill mechanism, the application of binding Markdown rules, and the promotion of Harness in practice, combined with the leap in the reasoning capabilities of underlying models, Agents have truly gained the ability to deliver production-grade results—not only faster than humans (much faster, in fact), but in many local areas even more rigorous than humans.

Take my own writing as an example. In the past, to write a good article, I needed to spend a lot of time conceptualizing the thread—clarifying the causes and consequences and the line of thought. For longer articles, I might first need to use pen and paper or a mind-mapping tool such as XMind to lay out the ideas and structure clearly before proceeding; I might need to look up some materials to substantiate my thoughts and viewpoints clearly and draw readers into the flow state I call the "trap of argumentation."

Now I basically never start from 0; instead, I act more as an architect defining the skeleton of the article, the core arguments, and the reasoning logic, listing points 1, 2, 3, 4, and leaving the remaining research, filling in of evidence, and logic gap-checking to the Agent. Although current AI writing occasionally still has a preference for sentence patterns like "not... but..." and a deliberate "human touch," it already conveys core ideas clearly and rigorously enough, and for many articles that is sufficient—especially since Gemini 3.7 Flash's writing remains very much on point recently, and its speed is terrifyingly fast; I can confidently hand almost most writing to it.

The evolution in programming scenarios is even more intuitive. Having used next-generation client tools such as Codex frequently since the beginning of the year, I have fully witnessed the alternating evolution of models and Harness engineering, and the experience is no longer what it once was. At the beginning of the year, when I introduced Codex to a developer friend, I said its capabilities were good, the code it wrote was also good, I was very satisfied, and it was worlds apart from Trae. He scoffed and said, "Code written by AI is simply unusable"—now, while I am writing articles, he is probably using his 20x Codex to directly vibe code, while scrolling through Douyin himself...

From writing code to market analysis, in almost all digital and knowledge workflows, AI has qualitatively transformed from "a junior who can get things running" into "a collaborator who can deliver."

Exploring the boundaries, but the bottleneck seems to be people.

AI still hallucinates and will cut corners when writing test cases in order to pass. If users themselves lack judgment, they will be easily blinded and deceived by it.

People often say, "AI is an amplifier of capability, not a wishing well." It can amplify your professional accumulation and breadth of thinking, but it cannot conceive for you blind spots you have never touched. In other words, the AI in everyone's hands is the same; the difference is the person—people are the boundary bottleneck of Agent capability,we used to describe insufficient execution ability as "high aspirations but low hands," but in the AI era,"low vision" is the most fatal bottleneck。

AI has learned knowledge that a person might not finish learning in two lifetimes, so from this point alone, saying that Agents have capability boundaries is a bit untenable. What truly gives Agents boundaries is, always, people.

A person who knows what excellent design, elegant architecture, or deep writing looks like can guide AI toward outstanding results even if they write code slowly; conversely, if you know nothing about system architecture and don't even have a concept of Git and version management, you simply cannot issue effective architectural instructions. This is not a matter of language organization ability—large models already have an extremely high tolerance for vague prompts—rather, the depth of your understanding of the essence of things determines how much room for action you can open up for the Agent.

Harness's own working environment is the most important

People often ask me: "Why can your Agent do so many things, while mine often can't even execute basic instructions?" The answer often lies not in the model itself, but in Harness. Simply put, Harness is the environment in which AI runs; you need to configure this environment for it so that it can execute tasks well.

This is still quite obscure, so let's use actual work as an analogy. Think of the Agent as a newly hired employee: just assigning work to them won't work; they need a computer with properly configured permissions, a network environment, clear code standards, business documentation, and collaboration mechanisms—people need these, and Agents need them too. Harness is what provides AI with such a working environment.

For an Agent, Harness means: an AI-friendly structured knowledge base (such as Markdown), plug-and-play Skills, MCP plugins called on demand, and clear task boundaries. Based on my own experience of stumbling through pitfalls, my experience can be broken down into the following points.

Accumulate deterministic knowledge and best practices

Things you've done, pitfalls you've stepped in, and paths you've successfully navigated are best written into documentation. You don't necessarily have to write this documentation yourself—you can have AI help you—but accumulation is essential. You can use tools like Feishu or DingTalk; once the CLI is configured, AI can read them. You can also directly use products like GitHub, using Markdown format, which is more friendly to AI. In short, it's important to let AI be able to read this explicit knowledge. You need to accumulate this yourself, because everyone's needs and the problems they encounter are different. Your dedicated context library is your core asset.

Package high-frequency repetitive tasks as Skills

Let repetitive operations be automated. Skills teach AI how to do this, and trigger them actively or passively (as you prefer) when you need them. But please note that Skills have extremely high personal timeliness and AI capability timeliness. It's possible that the capability you rely on will no longer be supported later, in which case you'll need to update your Skill. This is also why I manage them using soft links. In short, turn repetitive tasks into Skills and maintain them promptly.

Make good use of Plugins to make up for shortcomings in single-point capabilities

For tasks that involve cross-system calls, depend on specific environments, or require multiple capabilities to work together, provide stable external capability expansion by encapsulating MCP Plugins. Some mechanical, repetitive tasks cannot be handled by relying on Skills alone. For example, if a task needs MCP support and requires multiple Skills to collaborate, then you can encapsulate them as Plugins; they are better suited to this type of task that requires support from the other party.

Divide work among multiple Agents, and give up the obsession with "single end-to-end delivery"

Never expect a single Agent to do things perfectly in one go, especially for complex tasks. When facing complex systems, a single Agent can easily fall into context degradation and infinite loops.

Agents are the same as people. If you break down a task and assign it to different subordinates, and they divide the work and complete it, each of them will work very attentively, so every module will be done exquisitely and thoroughly. If you do it yourself, of course you can also complete it, but it may become relatively simple and shallow, not as in-depth. The same is true for Agents. Break down tasks and assign them to different specialized Agents, letting them dig deep within their own narrow and clear contexts, and finally converge and integrate. For engineering tasks, try using Worktree more often, just like how I previously had Agents collaborate inFeishuandOrca. Those interested can take a look.

Build an external memory layer

Contexts between different Agent systems are often fragmented from one another. Introduce unified external memory or a structured knowledge base to serve as a public context center for cross-tool collaboration.

Agents from different vendors all have independent memory systems, but they are isolated from one another and do not share. You can use some third-party tools to help you achieve a "cyber shared brain", such as EverMind. I have also written an article to introduce it before (Click to jump）。

Summary

In summary, Harness is very important. An Agent without well-developed Harness support is just a chat window that outputs probabilities; an Agent equipped with a complete infrastructure is a truly reliable digital employee.

AI is a tool; never cede "thinking" to AI

Many people fail to use AI well, essentially because they fall into "technological blind obedience"—they lack the ability to review generated results and hand over core business decisions to the model.

AI only helps you execute; it absolutely cannot help you decide, because the consequences of decisions are borne by humans, not by Codex on the computer or Claude Code. I recently noticed that many people do the exact opposite: they let AI think through problems for them and end up leaving execution to humans—this is putting the cart before the horse.

The most common pseudo-innovation is trying to build a "fully automated flow system": input rough market requirements, fully automatically generate code and deploy it, leaving humans only for formal approval at the final stage. This is essentially putting the cart before the horse—degrading humans into an execution layer that takes the blame, while pushing AI without logical verification up to the decision layer—this is not truly AI-native, but a deformed monster created to cater to the concept of AI, as absurd as Maurizio Cattelan's "Comedian."

Just likeBaoyuand an excellent manager I recently came into contact with believe, a highly efficient AI-native organization needs not tool people who passively review code, but a group of pathfinders with strong business understanding, high self-drive, and the courage to continuously trial and error at AI boundaries and converge experience. From another perspective: if a job can be perfectly delivered in a closed loop by AI without your business judgment, then the moat of that job itself is extremely weak, and it faces the risk of marginal costs dropping to zero at any time.

AI has not created new demand; what it reshapes is the service boundary

Looking back at the early rise of the Internet, technology often directly spawned entirely new user demands and business models, but the current evolution of AI is more about transforming delivery efficiency.

You can see marketing accounts telling you in a video just a dozen seconds long the narrative of "helping a company implement an AI system and quickly turn a profit," but this kind of thing has many logical holes that don't hold up to scrutiny—what scale of company is this? How many concurrent users? If it really is implemented, is it actually usable, or is it just for the boss to brag about? And even without AI, engineering teams of this kind have long existed. Enterprises' demand for cost reduction, efficiency gains, and automated development has existed for a long time; in the past, mature software outsourcing teams took on this work. AI did not conjure a new demand out of thin air; it merely lowered the delivery barrier on the supply side.

Then there is the currently heavily promoted one-person company (OPC) model. Most of the so-called OPCs that have actually succeeded are in fact multi-person companies, just with few people, or they are positions that were already mechanically repetitive.

I have seen a visa agent who started a business on his own, contacting clients to help them apply for visas, using the then-explosive OpenClaw to handle the entire process. To be honest, my 60-year-old aunt at home could figure it out herself (and she really did). There is no difficulty in this at all; it is all standard procedure, and the common blockers in it can be exhaustively enumerated. It could probably even be done with a traditional Workflow. The essence of making money has not shifted: the market and customer demand still come first. It is just that in the past it was limited by the ceiling on manpower, and now AI has flattened the tool barrier, allowing a single person to carry the service throughput that previously required several or even dozens of people, which is what made this business model work—that's all.

So do not believe that learning AI means you can make money. The real demand is not in the information cocoons of X or social media; it is in the mud—sitting in front of a computer scrolling Xiaohongshu and X will not help you find real demand. Go out and see more, and come into contact with people from different classes and industries, and only then can you find real demand—they may look very low-end, even hopelessly rustic, but in those seemingly traditional, process-heavy, even not-yet-digitized physical and traditional industries, understanding their pain points and delivering value with technology is far more meaningful than getting high on technical parameters by yourself.

AI brings efficiency, and also anxiety

As AI approaches or even surpasses the level of junior and mid-level personnel on specialized vertical tasks, the anxiety of "being replaced" inevitably spreads. This is no longer simply the sense of crisis created by marketing accounts; it is a concrete anxiety with traceable signs. If in 2023 it was nonsense, by 2026 everything has become traceable. Because, as I said above, AI and Agents really can work this year, and may even do it better than many people who specialize in this work.

But the real world is not a superhero movie. There is no omniscient and omnipotent Jarvis, and we are not Tony Stark. And even if we had Jarvis, we would not have Tony Stark's super intelligence and powerful ability to think in parallel across multiple threads, let alone his superhuman energy. Moreover, even someone as powerful as Jarvis needs Iron Man to command him, and today's AI, which is not as powerful as Iron Man, is even more so. Stripped of human intent, taste, and validation against real scenarios, AI is just a pile of weight matrices in a server. Only when connected to people with active awareness does it become an exponential lever.

My personal view is that there is no need for anxiety or internal friction. Take me, a B2B designer, as an example: a company may stop hiring new designers because of AI (at least it will not add new ones), but it is absolutely impossible for it to lay off all its designers, because a real B2B enterprise with a product of considerable user scale would never give up its deep grasp of business logic, user value, and overall experience.

Rather than falling into internal friction brought on by grand narratives, it is better to pick a real pain point and start working on it. Play with AI, look for fun, and if you get tired of playing, go travel, read books, watch movies, fall in love, go for a walk, and discover that the world is still beautiful.

Keep learning, stay curious

Here we are again at the clichéd and preachy part, and I am also trying my best to avoid falling into clichés and sounding sermonizing and doctrinal, but I find that no matter how I express it, it always circles back here, as if no matter which way you go, you arrive at the same destination—a feeling of "all roads lead to the same place."

In the process of sharing about AI myself, people often ask me: "How do I systematically learn Agents from scratch?" My answer is always very straightforward: "Spend 700 yuan to buy a Codex 5x Pro subscription, then whenever you run into a problem, throw it at it to do. Use it intensively for 1 month, and see whether you can or not."—That is exactly how I experienced it myself.

Moreover, many people always hold the mindset that as long as they go to Bilibili or some other media platform, find a course to watch, and take some notes, that counts as having learned it. Perhaps this is true for some purely theoretical knowledge, such as what AI is or what an Agent is, but I think that if you want to use AI well, this is all just the Zhao Kuo of the new era—armchair strategizing. At this point where tools are changing with each passing day, you cannot learn it just by watching video courses and organizing notes. Today's AI ecosystem is very much like the garage era at the birth of the personal computer: there is a lack of out-of-the-box silver bullets, and most of what exists is experience that engineers and geeks have figured out through trial and error amid fragmented software and hardware. A Skill that runs perfectly on your machine may, in a different environment, constantly throw errors due to missing dependencies.

Therefore, in this era, you need to keep learning and supplementing your knowledge, but perhaps it is a bit different from what you understand, because this is an era where even salespeople know that writing code is a faster way to help them complete customer onboarding. Execution has become a rigid necessity in this era. Because AI makes the formerly daunting "just try it first" very cheap (I will not describe it as low-cost), so when you are curious and simply unwilling to use AI because you think AI does not do a good job, please try it with AI first before saying anything, and do it first before deciding whether it is good or not.

So how should you do it? The key to breaking through lies in a closed loop driven by a clear goal:

• First, you need some motivation that drives you to do things, for example, you want to build your own personal website, even if it is just to show off (yes, that is exactly how I thought at the beginning), or you want to make a growth app for your family's kitten, and so on. They can be big or small—big enough to build a system or to earn you a high salary or even financial freedom, or small enough to be just a little tool for your own use, or simply to help you write articles—but the core is that this is a trigger, a reason, that can get you moving;

• Then you start executing. Strong execution will drive you to be more focused and more self-driven as you continue. In this process, you will inevitably encounter setbacks and failures—perhaps AI wrote the wrong code, perhaps the result is not what you wanted, and so on—but none of that matters. If your knowledge is insufficient, you can go learn. Open ChatGPT or Doubao and just ask; if you do not understand, keep asking until you can no longer come up with questions. Fast learning is very fast now. You can boldly let AI help you learn: if you do not understand, keep asking, have it give examples to help you understand, and so on. AI's ability in popular science far exceeds that of any human expert—it is so versatile and knowledgeable—this is what the old saying calls "learning by doing."

In short, in the AI era, those who are already willing to stay curious and have the drive to learn and explore on their own will do better; they are the elite of this era. Perhaps in the previous era they were also ordinary, but in this era their stage will come.

Using AI is joyful

In college, a teacher I greatly respected once said: "Doing design is a joyful thing." After half a year of intensive practice, I also want to say:Exploring AI is likewise an extremely joyful thing.

For the first time, it allows us to concretely present on screen, at extremely low marginal cost, those ideas once constrained by energy, coding ability, or tedious chores. That flow of regaining creative autonomy, and the sense of fulfillment from resonating with excellent fellow travelers amid the technological wave, are the deepest motivation sustaining our continued exploration. AI is not just efficiency; after all, life is long, and AI efficiency gains are only a small part of life. The joy AI brings is perhaps the real reason why everyone who explores and loves AI remains so captivated.