Installing an AI Harness

I'm a software engineer, working on multiple platforms employing user interfaces and robotic realtime systems. Each has code bases between about 50,000 and 500,000 lines of C++ code. At work, we now have access to Cursor, a "harness" which employs AI agents to solve problems. It automatically selects (or lets you select) the models needed for the task, for example Claude Opus, the latest ChatGPT version, Grok, DeepSeek, or whatever is available on the net. Cursor makes it a breeze to work through a code base and do whatever is needed to maintain it. And it's good! Not only at creating new code, but at maintaining existing large code bases, finding bugs after I describe the symptoms to it, and documenting software functionality for either the engineers working on it or the user. That last one really impressed me. I was tasked with adapting a user manual for a remote control interface from one of our platforms to another unrelated one. This would probably have cost me more than three months if I'd done this manually. It took a few runs with Cursor sifting through the code and then refining the results. The whole thing was done within a few weeks.

The one thing to be cautious about is how far you want it to go for you. It basically can do everything, including changing the code on its own and making a software build for you. At work, I limit it to propose patches on the output terminal, I don't let it change files. That way, I ensure that the philosophy behind the software stays in place - often it pays to be careful here, because Cursor might misunderstand what you intended. Yet, it's remarkable how well these models can adapt to your coding style and the ideas behind them.

At home, I have occasional need for writing software. For example, I wanted to look into reviewing my genealogy software, which I custom wrote to maintain overview of my ancestors back to 1500: when you want to do such a thing, then you create an incredibly dynamic data set, which needs to be properly documented in LaTeX and XML format, and which must be able to automatically draw a pedigree using GraphViz. There are constantly new members or even new generations joining the tree, or more information about specific members becomes available. This software works well, but I have extensions in mind, and something like Cursor can help. 

But Cursor costs money on top of the amounts you will have to pay vendors such as Anthropic, OpenAI, Amazon, Tesla, or whoever, for their services. There's free harness software called OpenCode, which I tried. It resembles Cursor and is easy to install. To be clear: using the models will still cost you money, but the harness doesn't add any fee. The harness handles the payments for you, you charge it up with $20, and it will then run until you exhausted the money. And it performs! I tried it by building a simple two-tone audio generator for my ham radio test equipment,. That audio generator was written and extended by features like saving and loading parameters to a file within about one hour. It's now not only giving me two fixed sine wave outputs, but it can make all sorts of silly SciFi sounds as well - maybe I'll build it into a robot for Halloween.... It would have taken myself a few days and quite some research. Yeah, I've vibe-coded it, and it was fun! Price: $0; I used the free "Big Pickle" model - slow but surprisingly strong.

I did that on Linux, but that really doesn't matter. You can do that on your preferred operating system.

Oh, one crucial remark: don't work on proprietary or privacy-critical material with it. The moment you uploade any code, it will become training material for these AIs. And who knows what else these companies do with the data you gift them.