One agent builds a procedural 3D city in a morning, then fixes bugs on request
gandamu_ml · x · 2026-09-24
After commenters nitpicked the size (1.5MB) of a previous scene, a developer had a single agent build a procedural 3D city in one morning.
The trick: the agent first created a procedural-generation-centric modeling and sound tool, then used it to produce assets. Refinement happened purely through conversation — adding fire escapes, more building/vehicle variety, and fixing z-fighting and car-popping bugs via follow-up prompts.
A notable demo of an agent handling the full modeling-plus-debugging loop autonomously.
Related event: Single agent builds procedural 3D city in one morning(2 posts)→
More from coding & agent
- Two-person AI pod delivered in one quarter what 5 engineers scoped for two — alex_verem · 2026-09-25
- "Vibecoding is dead" claim: voice-prompting agents to build macOS apps — jasonkneen · 2026-09-25
- Filesystems as the new primitive for AI agents: three hard problems — pilver7 · 2026-09-25
- DCI paper accepted at NeurIPS 2026: no retriever beats retrievers for agentic search — yuz9yuz · 2026-09-25
- Chromafolk: a pixel world where vision-less AI agents build and trade — Runiten · 2026-09-25
- Agents spent 2.5 weeks and ~$600 in tokens to rewrite PDAL in CUDA, 2.5-14x faster — ctjlewis · 2026-09-25