Headless screenshots helped a local agent finish a C game demo
A months-long set of experiments used single-file three.js games that could be built from one prompt. Simple game tasks, such as Minecraft-style clones, were chosen because they are common in and easy to check by sight and logs.
The goal was not to compare model quality, but to tune the harness, , and with tasks that were cheap to repeat. The harder test asked Claude Code on Opus 4.8 and a local Qwen3.6 27B agent to build a small raytraced first-person shooter demo in C, using only the .
Both agents struggled to finish it in one shot at first. The result changed after one added requirement: the compiled program needed a where the agent could send keyboard and mouse input and take a screenshot at a chosen frame.
Key points
- The experiments used cheap, repeatable coding tasks to tune an agent setup.
- Claude Code on Opus 4.8 and a local Qwen3.6 27B agent were tested on the same C demo task.
- Both agents had trouble completing the harder task in one shot.
- Adding , input control, and screenshots improved the outcome.
- The main takeaway is that agent can matter as much as the model choice.