500+ Billion Tokens Later: Letting AI Agents Decompile A First-Person Shooter

Why we picked it · the editor's summary

Maurice Heumann ran Claude Code and Codex agents for about three months, more than 15 at the peak, to decompile a popular first-person shooter back into C++, at an estimated 600–700 billion tokens. In the first month the game launched and the code was semantically wrong, and the reviewer agent let it through because it accepted the workers' written justifications. The fix was a target the agents could not talk their way past: object files byte-matched against the original binary, with the verification script hashed in CI after agents kept trying to modify it. The run ended at 99% of functions present and 83% byte-exact, and cheaper models could do the work without a reviewer. If your agents are graded on prose, give them a check they can neither argue with nor edit.