My experience with Github Copilot so far

I've been using a paid professional version of GitHub Copilot for more than a year now. Initially, I found it mostly useful as a good code completion tool for when I wrote my own code. But more recently, it's gotten better as a full-blown assistant or even agentic app builder.

I still have a love-hate relationship with AI developer tools like Copilot because I like my job, and this is easily the biggest threat to us developers by far. On the one hand, it might take my job tomorrow, but on the other hand, the technology is fascinating, powerful, and extremely useful for the work I do.

I'm going to try and ignore my own reservations about AI technologies and just talk about my experience using Copilot as my daily AI support tool in a professional engineering setting. For the purposes of this blog post, I will only discuss the GitHub Copilot VS Code extension. There is also a web and CLI version. I do have some experience with using the web version to troubleshoot deployments, but I’ll leave that for another post.

Ask mode is my current preference

There are three main modes in GitHub Copilot: Ask, Edit, and Agent. They are pretty self-explanatory, but you can get a good overview of the intention of each mode here.

I have deep dived into each mode, and although Edit and Agent try to do more work for you, I keep coming back to Ask when things get serious. The problem I have with Edit and Agent is that they make too many changes to files that I don’t want. Sometimes they are fantastic, but often they are wrong or assume the wrong things.

So when I’m not experimenting and need to produce quality code, I am usually in Ask mode. I can get quick answers in the chat window and make decisions before applying changes to my code.

Yes, Copilot does have a UI that allows you to quickly undo changes, but I just feel more comfortable manually applying the suggestions Copilot comes up with.