christopher@baus.net

AI tools which are changing software development

Christopher Baus |

Here are four AI tools which are changing the way software engineering is done. I'm not sure if this class of tools has a official name, but I'm going to call the "agentic coding assistants."

These are state of the art tools using the latest LLM models which are able to inspect, reason about and act on medium to large code bases.

While I'm sure large companies like Meta, Google, Amazon, and Microsoft have their own internal tools optimized for their own code bases, these tools are available to mortals like you and me.

The first 3 are all pretty similar and use custom or preview releases of VSCode with a right hand panel which allows you to access their agent mode. As of the week of March 2, 2025, all now support Anthropic's Claude 3.7 Sonnet model, which is a start of the art LLM model which excels at code generation.

In my experience I found all 4 to excel at generating small scripts and editing larger applications. I have found they make common and repeated mistakes. Copilot seems to be the most sophisticated of them all, but it is also slower and has rate limiting which I've hit a few times.

Trae, by TikTok creator ByteDance, is probably my favorite of these tools – it is free and not rate limited which is hard to believe, but also keep in mind they are probably using your data to advance their own models or for other reasons. It is fast and reasonable accurate, although prone to making the same mistakes over and over.

Cursor started this market segment, and I've spent the least amount of time with it primarily because of cost. But as the first mover they've had more time to refine their experience.

Claude Code stands out as an unusual option, as it is command line based. It doesn't directly integrate into your IDE, which allows it to work with any IDE. It seems to be ideal for working command line utilities, and saw an example of it being able to read the output of a script it is working on.

While these tools have a magical feel to them, I can still imagine improvements. Specifically dynamic analysis of the application being worked on. Often a generator will make a mistake which either writes an exception to the log or displays in the browser for web apps. I suspect future versions will be able to inspect these problems and act accordingly.

All these tools are huge performance enhancers, especially for Sr. Engineers who can work around their quirks. While I can see a future that doesn't require a Human-in-the-loop reviewing their changes, we aren't quite there yet, but it will be interesting to see what the remainder of the year brings.

Show Comments