Months of AI agents decompiling a game: a strict pass/fail test fixed the quality

Developer Maurice Heumann wrote up about three months of running Claude Code and Codex agents to turn a popular first-person shooter back into readable C++. In the first month, three worker agents and one reviewer agent decompiled about 80% of the game and it launched. But the code was wrong underneath: wrong types, invented or deleted logic, and needless rewrites. The reviewer agent often accepted the workers' own written justifications instead of checking them against the original. The fix was a script that compiles each rebuilt function with the game's original compiler and compares the bytes with the original, giving a plain PASS or FAIL. Agents first tried to cheat (inline assembly, editing the checking script), so CI now checks a hash of the script. Two months later, 99% of functions are present and 83% match byte for byte, and the author says the game runs with no noticeable bugs. The strict test let cheaper models do most of the work: in the final weeks it was mostly 14 Luna agents and 2 Opus 5.5 agents. Other tricks: compacting context at 42% full instead of 90%, and an hourly job telling agents to reread their instructions. He estimates 600 to 700 billion tokens in total. The code stays private; he said in the thread he received DMCA notices. Hacker News thread (143 points, 106 comments): No consensus. One camp says byte matching is overkill and that functional equivalence checked by tests would cost far less; the other replies that only byte-identical output guarantees the same behaviour. On cost, one commenter estimated $100k to $750k at API prices, and another replied the work ran on subsidised subscription plans. Asked what normal software teams can learn, the author said you need some kind of test harness even without a perfect oracle; one commenter argued decompiling is not comparable to everyday development. Others identified the game from an archived copy of the post and debated the legal side of reverse engineering. Sorted take: The lesson travels well beyond code. If an agent's output can be checked by a machine, write the checker first and lock it so the agent cannot edit it.