0/140k LOC - Using agents to build complex apps that actually work

Everyone has the "agentic coding best practices" talk: prompt like this, use subagents, write a good context file. But look at who's actually shipping complex software entirely with agents — it's either enterprise teams with armies, or it's todo-app demos. The interesting middle is mostly empty.

This talk is a field report from that middle. SpriteMachine is a professional pixel-art editor — a custom WebGPU renderer, real-time pointer and pressure pipeline, deterministic undo/redo, a sandboxed scripting engine, dual web/desktop deployment. ~140,000 lines of TypeScript and 45 hand-tuned GPU shaders. Not one line of source was written by hand.

I'll argue the bottleneck on agentic coding isn't the model — it's whether your codebase can tell the agent it's wrong before you have to. We'll walk through the concrete machinery that makes an all-agent codebase hold together instead of rotting.

None of it is AI-specific technology. What's new is the reason: your fastest contributor now has zero judgment, so legibility and enforcement stop being nice-to-haves. Build a codebase that argues back.

Key takeaway

Shipping complex apps with agents is an architecture-and-enforcement problem, not a prompting problem — and you can apply all of it in any stack on Monday.