I have been using an AI tool to upgrade this website.

I started by formulating a multi-step plan, then prompting the agent on each goal of each step.

Technically the agent - which in this case was MiniMax M2.7 - is very competent. I try to be hands off most of the time.

One technical issue it kept facing was when it tried to use @TemplateExtension very often, which are methods in Quarkus that expand any static non-private class's methods to be accessible template wide without any bindings. [Quarkus TemplateExtension] https://quarkus.io/guides/qute-reference#template_extension_methods) This caused multiple rollbacks, since many of the services were not static to begin with.

I also prompted it to remove some hard-coded headers and replace them with Forwarded headers, but it decided to create its own implementation rather than using Quarkus's own implementation. The agent also likes to recommend lots of charges but when it came time to implement them ,it discovers that it's already implemented. Maybe this is due to the context window.

I don't have any new project to ask it to do from scratch since I am still learning myself, but nudging the model to the right answer seems far better but very tedious reviewing and testing, that is. Speaking of tests, this might be on me. In my project there isn't any tests setup.

I am using Opencode as a coding agent. While it is easy to set up, it is very heavy on the CPU.

I will try to continue using this coding agent. Maybe I could try to set up a test suite using it.