1.2B tokens later: multi-agent coding run rebuilds a 3D game but hits diminishing returns
teortaxesTex · x · 2026-10-29
A blogger ran a 4-subagent coding team (Rust physics, WebGL2 rendering, WebAudio, evidence recovery) to rebuild a classic marble-roller 3D game, burning 1.2B tokens ($11.62 at peak API rates). V4.1F-1017 still shows classic '2025 LLM' failures — the first run coded itself into diminishing returns and a retry spent another 500M tokens mostly recovering prior state, with quaternion-handedness and physics-regression bugs dominating the transcript. Author's verdict: Claude/GPT would likely have been faster and maybe cheaper.
More from coding & agent
- OpenLolo open-sourced: a DIY hardware box that lets AI operate your iPhone via MCP, no jailbreak — Eonminion · 2026-10-29
- Dev permanently locked out of GitHub account by SMS 2FA, and GitHub won't restore access — m_schuetz · 2026-10-29
- $5 in evals cut his AI bill 74%: Claude Haiku 5.5 beat Gemini 3.8 Flash at 1/4 the price — MicahBerkley · 2026-10-29
- Running remote Ollama+Qwen3.8 on RunPod via Codex with just an API key — huangyun_122 · 2026-10-29
- How I turned Claude into a full ad agency with two connectors — one afternoon vs 3-6 weeks — PrajwalTomar_ · 2026-10-29
- How we share MCP servers company-wide without giving every agent every credential — Kluff_AI · 2026-10-29