Original video in Chinese.
Key Takeaway
- Replit Agent is an aggressive AI coding tool that uses an Agent to handle environment setup, code writing, and project deployment, enabling cloud-based development.
- Replit Agent works like a “catering service team”: the user only needs to provide the requirements, and the Agent takes care of the actual execution, making it suitable for rapid prototyping.
- Compared with tools like Cursor, Replit Agent puts more emphasis on automation and reduces human involvement.
- Replit Agent supports development of simple projects such as H5 pages, but it still has limitations for complex projects.
- The article emphasizes that AI coding tools can greatly improve efficiency, but users still need development thinking. AI is currently in a stage where “it gets stronger when it meets the strong, and weaker when it meets the weak.”
Full Content
Large language models have been hot for almost two years, and I didn’t expect AI coding to be the first area to really take off.
The newly released Replit Agent is especially aggressive. From environment setup, to code writing, to project deployment, it uses an Agent to handle everything. During the process, it will sometimes ask for your opinion, such as whether this UI matches what you expected. Most of the time, Replit Agent is just quietly working.
The whole development process is completed in the cloud, using its environment, which is quite reassuring. More importantly, this means you can use it with any device you want, even a phone. Once you have an idea, you can quickly finish a prototype. Then you can download the entire project as a ZIP file and do further work locally.
This mode of Replit Agent reminds me of the catering service teams in the countryside, the kind that specialize in weddings and funerals. They provide the chef, kitchenware, tableware, tables and chairs, and even the ingredients and seasonings. You just show up and eat.
By comparison, Cursor is a human-machine co-work model, and still requires a lot of human involvement. And the earlier GitHub Copilot is even more behind the times; I think it can only be considered intelligent code completion.
To test it, I used Replit to make a simple H5 page. After the user enters an API Key, they can choose an OpenAI model and chat.
After the requirements were sent over, Replit first did an analysis, and based on its understanding, it also gave me some options to see whether I wanted to flesh those out as well. This step is essentially requirement confirmation. Anyone who has worked with clients knows that if the requirements aren’t clear, this project is very likely to be painful.
Once it got started, Replit automatically deployed the environment, downloaded and installed all the required packages, then wrote the code, configured the database, and so on.
Because the project was relatively simple, the testing process went very smoothly. Calling the OpenAI model for a chat worked on the first try. The rest were just small fixes, like removing unnecessary elements from the UI; I only had to say it once and it was done.
Replit can’t speak Chinese, but it can still understand you if you use Chinese. Still, I recommend using English to avoid ambiguity from translating Chinese into English. English at around a fourth-grade level is enough to communicate with the AI; it understands it all.
When it came to deployment, there wasn’t really anything to adjust. Replit had already set everything up for you; you just click Start and that’s it. Once the progress bar reaches the end, Replit provides a link, and that is our H5 page. I tried it on desktop, and it worked perfectly normally. Opening it on a phone also worked.
Replit supports auto-scale, which means if traffic gets heavy, it will automatically upgrade the configuration to keep the project running. You can see analytics in the backend. If you have your own domain, you can also bind it in the backend.
By now you should understand why I said Replit is like a rural catering service team. But right now this team still isn’t very capable. It has no problem handling small jobs, like H5 pages, but once it runs into a more complex project, it just can’t handle it.
The day before, I had Replit make a WeChat mini program too, also a chatbot connected to the OpenAI API. The whole process was really frustrating. Replit kept asking me to check this and check that, and even asked me to check whether there was something wrong with the Python documentation—who’s working for whom here? When it came to deployment, the very last step absolutely would not go through, and I couldn’t figure out why, so it just got stuck there.
So I estimate Replit will probably need at least another six months of iteration. For now I’m not subscribing; I’d rather keep using Cursor, which is more reliable.
Let me show you: I built my personal blog with Cursor. The text transcripts for these videos, as well as my daily reading list, will all be published on the blog. Going forward, I plan to use Cursor to build a chatbot, add RAG, connect it to my personal knowledge base, and open it up for everyone to use.
Among these AI coding tools right now, I’ve found two camps of opinion: one says these tools are all trash, and the other says they can replace programmers. I disagree with both. I think AI is currently in the middle ground: it can’t replace people, but it can greatly improve efficiency. If you’re a complete beginner, it’s really hard; at the very least, you need to have development thinking.
Just like I said in one of my videos earlier this year, the characteristic of AI at this stage can be summed up in eight words:
It gets stronger when it meets the strong, and weaker when it meets the weak.
AI isn’t that magical. Right now it’s just a Copilot, and it needs a very strong Pilot to make it useful. If you’re not that strong, then just wait—wait until the day it becomes widespread, wait until the day it gets very close to AGI, and you can benefit along with everyone else.
OK, that’s all for this episode. If you want to find me, come to the newtype community. See you next time!