One person, 588 agents, F-Zero X on a 3DS
One developer brought the racing game F-Zero X to the Nintendo 3DS in 11 weeks without opening an editor. Claude agents wrote all the code; the human set goals, played the builds and approved every merge.
ContextF-Zero X is a 1998 Nintendo 64 racer. Fans rebuilt its source code from the game (a decompilation), and others ported that to PC; this project took the PC port to Nintendo's 3DS handheld.
The project, called G-Diffuser, ran from August 11 to October 26. One long-lived Claude Code session acted as orchestrator: it held the project memory, wrote briefs for background agents, merged their branches and was the only voice the human heard. The postmortem sums up the human's role as "a captain, not a programmer".
The postmortem counts 777 human prompts, 588 subagents, 1,288 workflow agents and about 22.6 billion tokens, which the author estimates at $19,559 at API list prices. On a New 3DS, every one-second window of a 3D race measured between 59 and 61 fps; the older, slower 3DS reaches a 60 fps median.
The speed came from a renderer that writes the 3DS graphics chip's own commands from data prepared ahead of time, instead of translating Nintendo 64 draw calls every frame. Hardware counters showed the old 3DS was starved by instruction-cache misses, so race loops were resized to fit its 16 KB cache. A renderer that won in the emulator lost on hardware and was dropped.
588subagents, steered through 777 human prompts
Technical details
Releases: v0.9.0 on September 4 (New 3DS only); v1.0.0 on October 19 (old 3DS support, the native renderer on by default, the 64DD Expansion Kit on New 3DS, and a bottom-screen dashboard called DASH); v1.1.0 on October 26 (2 to 4 player multiplayer, and both 3D eyes drawn in one GPU submission). No game data ships; players extract their own.
Old 3DS: the first boot on hardware ran at about 10 fps. Now 85% of one-second windows in a 3D race sit at 60, with a 10th percentile of 56, though the heaviest courses can still fall to 30. Drawing both eyes in one submission lifted the 3D menu from 40.2 to 44.6 fps. The port layer is about 181,000 lines under port/3ds, tests included.
Shelved along the way: a GPU vertex-transform attempt, a working pose-replay fallback (the human chose to aim for a real 60 instead), and a retained renderer that only won in the emulator. The Expansion Kit stayed New 3DS only because it was too slow on the old model.
Cost: the $19,559 is the author's estimate at list prices; the page also logs $15,567 in billing cache writes. The orchestrator's memory file reached 591 lines.
Why it mattersThis is one of the most detailed public records of a long, agent-built project: what the human did, what the agents did, what it cost and what broke. What transfers is the discipline around the agents, not the model.
For youUE plugin work has the same shape: an orchestrator session with a memory file, a worktree subagent per feature, and a killswitch plus a log receipt for each change. Copy the rules in the next block into your subagent brief template.
- decompilation
- Source code rebuilt by hand from a finished game's machine code, so it can be changed and recompiled.
- orchestrator
- The one agent session that hands out work to other agents, keeps the plan and merges what they return.
- subagents
- Extra AI agents started for one task each, which report back to the agent that started them.
- instruction-cache
- A tiny, fast memory on the processor that holds the code it is about to run; code that does not fit has to be fetched slowly.
- emulator
- A program that imitates a console on a PC, handy for testing but not always faithful to real hardware speed.
Sources G-Diffuser postmortemPostmortem: the agentic workflowRepository
