The Impact of Vibe Coding on Individuals and Organizations | Liao Jiayi

The Impact of Vibe Coding on Individuals and Organizations

Liao Jiayi Liao Jiayi

Two years into vibe coding, AI agents have already changed how individuals work and made it urgent for organizations to change how they operate.

Translated from Chinese with AI · Read the original

I want to talk about vibe coding today. It’s come a long way in the past two years. Looking back, all those gains in coding speed eventually added up to something bigger. They changed how I think about code and products altogether.

1. Introduction — How my development habits changed

The funniest thing about the AI era is that every two or three months, you have to learn a new tool because it can make you two or three times as productive. That’s a complete reversal in how I think about tools. Five years ago, I would’ve said getting good at your tools mattered more than which ones you used. Awkward, lol.

Here’s how my own programming workflow changed over that time.

When How I wrote code
Early 2025 Manus and YouWare started taking off. I saw people all over social media sharing what they’d made, and coding started to feel like something anyone could do. I tried a few of these tools, mostly for visualizations and personal blogs. Building something had gotten easier, but maintaining and improving it over time was still a problem.
First half of 2025 I started writing code with ChatGPT. For a small task, I’d throw it over to ChatGPT, paste the code into my IDE, and run it. I’d fix minor problems myself and paste bigger ones back into the chat. After a few rounds of back-and-forth, the code usually ran the way I wanted.
Second half of 2025 I started using Trae, ByteDance’s internal IDE, at work. A lot of the Chinese models I tried kept running into problems, and our internal systems weren’t integrated with AI. I still spent a lot of time looking up docs and code examples, then pasting them into Trae so it could pick up from there. It was barely usable and honestly pretty frustrating.
Late 2025 I switched to Cursor and immediately uninstalled Visual Studio. Cursor felt like the first truly AI-native coding tool I’d used. My favorite feature was that it could read my terminal directly and fix its own code based on the errors a command returned.
Early 2026 Once I tried Opus 4.5, I felt Claude Code had a better harness than Cursor. The model and the harness seemed to make each other better with each iteration. I reinstalled Visual Studio and started using it alongside the Claude Code CLI.
Mid-2026 Claude Desktop kept getting better, and after switching to Opus 4.7, I spent less and less time reading code. At first I only looked at the key files. For a lot of tasks, I eventually let AI commit and merge the changes on its own. Claude Desktop felt like all I needed.
Second half of 2026 Codex and Claude took turns pulling ahead, and I switched between the two apps. Codex has a less geeky interface, which plenty of people like. When GPT performed better, I used Codex. When Claude pulled ahead, I went back to Claude Desktop.
Now Claude Desktop is still my first choice. To me, Claude’s best model still has a slight edge over GPT’s best. But my Codex quota resets fairly often, so when I’m actually getting work done, I use whichever one still has credits.

Eight changes to the way I code in two years. I wouldn’t have imagined that before.

2. Where are the bottlenecks once you can work faster?

2.1 Running out of ideas

This is probably the first problem you’ll run into. When you’re used to traditional software development, one of the biggest shifts is going from finishing one thing in two months to finishing it in two days. On day one, you already need to be thinking about what comes next. Otherwise you’ll finish on day two and have no idea what to work on when day three rolls around. Running out of ideas is one of the biggest bottlenecks, and I think it’s the easiest one to hit. If you don’t have an idea, it hardly matters how capable the AI is.

2.2 Keeping up with context switching

Once you have enough ideas, you’ll probably be doing three or four things in parallel. Plenty of geeks are already juggling eight or nine. Can your brain switch between those contexts that quickly?

With five agents working for you, chances are they’ll break your day into little pieces. Sure, you can have agents delegate to other agents. You still have just as many decisions to make, and the quality of those decisions matters just as much. The agents’ work still has to follow the direction you set, and keeping them aligned takes effort. One bad call can easily send an agent off track.

Agents can also get things wrong, and every mistake can create new problems that need more decisions. It feels a lot like managing people in an established organization. You spend your day making all kinds of calls. Some are easy. Some are hard. Sometimes you make the call, the work goes sideways anyway, and you spend even more time putting it right.

All of this means constantly switching contexts. When the AI agents you’re managing can execute 100 times faster, they can come back for another decision at any moment. To manage that many agents, you have to switch contexts quickly and keep making good decisions.

2.3 Learning too slowly, with too much still to learn

I ask myself this every day. How can I learn things faster?

I think the AI era calls for getting hands-on and learning across a wider range of things. If AI makes you 10 or even 100 times as productive and you spend all that extra capacity doing what you’re already good at, you won’t grow. Making something new takes new skills. The exciting part is bringing what you’ve learned over the years together with what AI can do now and building something you couldn’t have made before.

Can we take in new ideas and learn as quickly as AI can execute? I think we’re still a long way off.

3. What should engineers do?

I imagine a lot of engineers are wondering the same thing. Now that we can code this fast, what are we supposed to do with all that time?

I think there are two paths for engineers as coding gets easier. You can go narrower and deeper, or you can broaden your scope. Those two career paths have always existed. I think the difference between them will get more pronounced.

3.1 A new mindset — do to learn

I came across this phrase on X. We used to learn so we could do things. Now we can do things to learn. That’s a big shift, especially for engineers who are eager to get their hands dirty. They can pick up the dry fundamentals while they’re actually building something.

There’s a trap here, though. If you keep practicing the same thing in the same field, you’ll keep drawing on what you already know. I’d practice across domains and try different kinds of work. That way, each project exposes you to something new and gives you a chance to grow.

3.2 Path one — Run the infrastructure AI depends on

The first path is to go into infrastructure and become the person who keeps AI systems running.

It sounds like dirty work. If AI can write the code and do anything you can do from a CLI, what’s left for you to do?

Say your boss asks you to provision a batch of machines. You need to talk to a supplier or, at the very least, click the button to provision them. You have to check the hardware models, negotiate how long you’ll have the machines, and explain what the business needs. Once the machines are ready, agents can figure out how to use them and which workloads should run where.

Give an agent access to a Kubernetes cluster and it can inspect all the machines and figure out how to match the setup to the workload. It can write the application code itself, too. Your work is mostly where AI meets the outside world. You’re on call, and you’re fielding questions from product, operations, and your boss. It sounds a bit like dirty work, but I don’t see AI taking it over anytime soon. If AI keeps improving, the people willing to do that work might be the ones who come out ahead.

3.3 Path two — Learn to lead the business side

The second path is to get directly involved in the business.

There used to be a middle ground where you could become an architect or a very senior engineer. I think there’s much less room there now. Even at the senior end, the coding itself is becoming easier for AI to take over. AI has also made architecture and implementation on complex projects far less time-consuming.

That frees you up to pay more attention to what engineering produces, which is the product. The product is part of the business, and you now have the time to work with the people running it. You can connect engineering and product and see the whole thing through end to end.

That brings us back to the earlier bottleneck. Do you have enough ideas that you can quickly build and validate end to end? If you don’t, most of that extra time will sit idle, with little to show for it.

4. How can organizations get the most out of this?

4.1 Rebuild your infrastructure around agents

This is the first thing every organization needs to do, and it needs to happen quickly. That might involve reducing headcount, restructuring, or taking other steps. Every day you put off making your infrastructure work with AI, the productivity gap gets wider. It’s like riding a regular train while everyone else has already switched to high-speed rail.

4.2 Make room for exceptional individuals

Steve Jobs once said the gap between a good engineer and a bad engineer was about 10x. I think it’s closer to 1,000x now. In any organization, especially a larger one, most people perform somewhere around the average. When someone exceptional comes along, the organization should make room for them and give them more latitude. That includes a bigger token budget, more say in decisions, and a setup where they can actually use both.

In the past, we didn’t like people standing out too much because they still had to work with everyone else. Now there are fewer layers of coordination. Once an organization has rebuilt how it works around AI, the AI should in principle be able to run experiments without anyone else getting involved. That’s a huge advantage for someone who can see a project through end to end. As soon as they have an idea, they can start testing it.

I think someone like that can have an increasingly outsized impact, enough to determine the direction of a company. I also think it’s one reason so many talented people are starting companies early in their careers.

4.3 Organize teams around the business

ByteDance is already doing this. The Toutiao teams have brought frontend, backend, and other engineering functions together under business units and moved away from departments split by engineering specialty. I think that’s exactly the right call. Engineers should be able to work across platforms and, at a minimum, own a product’s development end to end. With AI, the stack no longer has to dictate who does the work. One person can cover iOS, Android, web, and the server side. If several people share the work, each should be able to cover for the others.

This cuts a lot of coordination overhead. In the past, just getting engineering departments to work together could mean bringing in a project management office and adding more people to handle coordination. The organization kept getting more bloated. With AI, we can move away from structures that need so many people just to keep the work moving.