AI agent writes x86_64 assembly and eBPF code to render fractals on /dev/fb0 in a browser-based Linux VM

An AI agent using a local Gemma model writes pure x86_64 assembly for Rule 30, Mandelbrot, and Julia fractals, then executes the code directly on a real Linux framebuffer device (/dev/fb0) running inside a virtualized browser environment (islo).

The setup uses a real VM with real kernel access instead of a container, enabling eBPF tracing via bpftrace, direct framebuffer access, and nested QEMU virtualization with KVM — capabilities blocked in standard container sandboxes.

The agent iterates through plan, build, review, and refine steps, with an oracle assembling each iteration, running it, and measuring both binary size and pixel accuracy against a Python reference; the final eBPF trace streams every syscall and framebuffer write alongside the live fractal render.

New islo.dev accounts receive $50 of free credit with no card required, sufficient to run the full convergence loop end-to-end.