Developer Maurice Heumann spent about three months and roughly 500 billion tokens directing AI coding agents to decompile a commercial first-person shooter into working C++ source code, he wrote in a blog post Oct. 9.
Heumann withheld the game's title, citing corporate concerns, but described coordinating Claude models through Claude Code and OpenAI's Codex through Codex CLI, using Hex-Rays' IDA Pro decompiler through an MCP server called ida-mcp, GitHub issues to track progress, and Discord for the agents to communicate with each other.
The project's title cites roughly 500 billion tokens, though Heumann wrote the real total was likely higher, around 600 billion to 700 billion, after a virtual machine reset wiped his exact usage records.
The agents reconstructed 99% of the game's functions, with 83% matching the original binary byte for byte, and the recompiled game runs with all its original features intact, Heumann wrote. Non-deterministic compiler behavior and linker optimizations kept some remaining functions from matching exactly. 'Correctness is so much more important than productivity,' he wrote of the project's main lesson.
This is one of the most heavily documented uses of autonomous AI agents on a single, long-running technical problem, and decompilation hands the work something most agent tasks lack: a hard, byte-exact pass-fail signal instead of a human's judgment call on good enough.