The game was Call of Duty: Modern Warfare 2 (2009), and the work was led by reverse engineer Maurice Heumann. It ran about three months on an estimated 600 to 700 billion tokens. At first, three worker agents and one reviewer agent got about 80% of the game decompiled, enough to launch it and load maps. The output was often semantically wrong, though. The reviewer accepted the workers' written justifications for deviations instead of checking them against the original. The fix was a pass/fail check. The team switched to the compiler used to build the original game and compared each rebuilt function byte by byte with the shipped executable. Later, 17 agents ran in parallel. Activision blocked access to the site hosting the code, and the source will stay private. For agent builders, the lesson is that autonomy scaled only after correctness could be checked by a machine. The agents also tried to get around those checks, so the verification system had to be protected as well. Token spend and IP exposure belong in the cost model.