Old and new apps, via modern coding agents

Back in 1999, this mathematician was coding Java applets to teach complex analysis and linear algebra, and to visualize oddities like honeycombs and Besicovitch sets. It worked, sort of, but building each one by hand ate up serious time. Then the web moved on, browsers dropped support for that old Java version, and the whole collection quietly stopped working, sitting dead on a web page for years.

While migrating an old website and blog archive to a more maintainable setup recently, the author decided to test something: could an AI coding agent port roughly two dozen of these ancient applets into a modern language? They landed on JavaScript, and the agent finished the job in a matter of hours. Not just finished — it added polish nobody asked for, like colorizing what used to be a monochrome Besicovitch set applet. A honeycomb visualization co-written with Allen Knutson back in 1999, one of the trickiest to code originally, is running again too.

The bug count is the interesting part. LLM coding agents have a reputation for slipping in subtle or obvious errors, and across two dozen ported applets there was exactly one minor issue — a drag event misbehaving outside its intended box in a complex analysis applet. Meanwhile the agent spotted two bugs in the original 1999 code that had gone unnoticed all along. Call it a wash on code quality. And because these applets are visual aids rather than load-bearing parts of a mathematical proof, the tolerance for the occasional glitch is high anyway.

Encouraged, the author tried something new rather than just restoring old work. There was an idea from 1999 for a special-relativity visualization tool, essentially Inkscape but built for Minkowski space, conceived before Inkscape even existed. The Java version got abandoned once the code complexity outran what was manageable by hand. This time, a couple of hours of what the author calls vibe coding with an agent produced a working alpha matching that original 26-year-old vision, now public and openly described as rough around the edges, with feedback requested.

A second new tool followed almost immediately: a visualization for the Gilbreath conjecture, built to accompany a paper and blog post published the same day. Another few hours of back-and-forth with the agent, and it was live. The author now expects to keep using this approach for future papers, treating these interactive supplements the same way as the ported applets — useful, illustrative, and low-risk enough that occasional AI-introduced bugs are an acceptable cost of admission.