Agent-Oriented Development: AI Coding Beyond Chatbots
Last time sharing collection of GPT-oriented development examples was in January
Most developments are small in scale,
the main form is to describe the requirements in the chatbot interface, then a person "moves" the code output by AI into the project, the person does testing, sends the results to GPT, and then GPT outputs the modified product. This cycle repeats ...
Now, GPT-oriented development is Agent-oriented development.
Agents can operate external environments through API and token, such as cloudflare, github, etc.
Agents can operate the chromium browser through tool_call. Not to mention bash terminal operations.
Agents have strong image recognition ability; they can understand GUI interfaces. You can let them start VNC, and you connect to it, so you see the same interface, making communication easier.
In this way, compared with chatbot-oriented development, the biggest advantage of Agent-oriented development is that the Agent can test the final product by itself, greatly saving the time humans previously spent on testing.
In particular, when you're not using a SOTA model, it's likely you cannot succeed in one shot development, so letting the Agent test and iterate in a loop is especially important.
Here is my collection of Agent-oriented development https://zelikk.blogspot.com/search/label/Agent开发
It's worth noting that none of these use SOTA models. They all use mimo, longcat, hy3, deepseek-v4-flash, etc., either picked up on the roadside that nobody else wanted, or particularly cheap ones. I hope to serve as an example to let readers see that for requirement development at this difficulty level, there's no need to use the most expensive models, so there won't be such a high cost, and more people are welcome to try to become creators, satisfying their own needs through development.
Another point I hope readers will see is what kind of expression to use when communicating with the Agent in Agent-oriented development. Furthermore, in your own brain, what kind of thinking mode to use.
For projects where the front-end is HTML, the Agent can debug through the chromium browser.
E.g.,
For projects that require logging into an account, if you're not comfortable giving passwords and other information to GPT, you can have the Agent start VNC, and after you connect to VNC, you complete the login operation, then let the Agent take over and continue debugging.
E.g.,
When considering the cost of cross-platform development, the front-end can use HTML, and the back-end can use GO. GO can be cross-compiled very conveniently; the Agent develops in a Linux environment, and the compiled result can be used in your own win10 environment.
E.g.,
Agents now have relatively strong image recognition ability, and c# language GUI development can also be directed through screenshots + markers.
E.g.,
After you have the Agent do something, you can consider asking it to summarize into a script, so it's convenient for you to reuse it later.
E.g.,
Posts
