HalfEdge

HalfEdge is an experimental attempt to develop an optimized Unity C# package with the help of coding agents like Claude Code and Gemini CLI.

For this experiment, I chose to implement a Half-Edge data structure library, inspired by wblut’s HE_Mesh. I chose it as a subject because it's practical, strictly unit-testable, and offers plenty of room for optimization — ideal traits for this kind of experiment.

How It Goes

First, I wrote a design document (with ChatGPT, of course), set up a Unity project with a unit test environment, and asked Claude Code to implement the library with unit tests. It generated several components fluently — impressive at first glance.

However, issues quickly became apparent. It produced many inverted faces, the Chamfer Edges modifier generated broken edges, and so on. These defects were obvious to the human eye but difficult to catch with automated unit tests. You can ask the agent to write specific unit tests, and it sometimes succeeds — but other times, it adds meaningless ones.

After resolving a number of minor bugs, the initial implementation (HalfEdgeMesh) was done. I then tried to optimize it, but the changes required were too extensive for an incremental approach. Also, longer tasks tend to fail more easily — especially with more affordable models like Claude Sonnet. So I pivoted to building a new version from scratch: HalfEdgeMesh2.

In HalfEdgeMesh2, I adopted an "optimize-first" strategy. I started by writing a new design document and then implemented only the core modules based on it. I had the agent write them in an already optimized form. This included zero GC allocation, Burst compiler support, custom collection classes, and the Unity Jobs System. With some human supervision, we successfully produced a suite of optimized, zero-GC C# components for Unity. Yay.

Next, I started porting over the generator and modifier components. But that's when my motivation dropped. These parts brought new issues and still needed my input to optimize properly. While it's easier than writing everything myself, I now clearly saw the time and effort involved — and it’s definitely not "free". So I paused the project there.

Lessons Learned

It’s possible to build a well-optimized Unity package using coding agents. You can achieve zero GC allocation, Burst optimization, parallel execution, and solid unit test coverage. But it still requires proper human supervision. It’s far easier with the agent’s help — but it’s not fully automatic.