IT Home, October 13 update: On October 9, developer Maurice Heumann released a project report stating that his team spent three months and consumed over 500 billion tokens to invoke AI agents to decompile the 2009 game "Call of Duty: Modern Warfare 2" and reconstruct it into C++ source code.
The project is led by Hoyman, with assistance from RektInator, Future, st0rm, and other community members.
For AI invocation, the team initially used both Claude Max (20x) and Codex Pro subscriptions, running AI agents through Claude Code CLI and Codex CLI. The models included Sonnet 5, Opus 5.5, Luna, Sol, and Terra, with agents tracking tasks via GitHub Issues and exchanging information through Discord.
In the first month of the project, four agents were deployed: three responsible for decompiling and submitting code, and one for review. By this stage, approximately 80% of the functions were completed, allowing the game to launch, display the main menu, and load maps; however, the team found that while the code was readable, it contained semantic errors at the level of function signatures, types, structure layouts, and logic.
Subsequently, the team introduced a "byte-by-byte matching decompilation" verification mechanism. This mechanism reconstructs the target file using the original game compiler and performs a byte-by-byte comparison between the reconstructed functions and those in the original EXE/PDB files; functions that pass are marked as PASS, while those that fail are revised by the agent.
To prevent agents from bypassing validation, the team prohibits inline assembly, naked functions, object patching, and embedding bytes directly in code, and verifies script hashes against GitHub Actions secrets via CI. This mechanism enables low-cost models to perform high-accuracy tasks under strict feedback, replacing manual or agent-based review processes.
In the later stage of the project, 14 Luna agents and 2 Opus 5.5 agents primarily operated, submitting code via independent branches and pull requests. Ultimately, 99% of the game functions were incorporated into the reconstructed source code, and 83% achieved byte-for-byte matching; the remaining functions could not be fully matched due to non-deterministic characteristics or linker behavior.
Byte-for-byte matching refers to comparing the binary data of a reconstructed function, after compilation, with the corresponding function in the original program on a byte-by-byte basis. If the data matches exactly, it typically indicates that the function retains the same semantics and behavior in the target compilation environment as in the original.
Hoymann stated that the relevant code will not be made public, and the reconstruction results will be retained solely for private use.
IT Home provides the reference link: 500+ Billion Tokens Later: Letting AI Agents Decompile A First-Person Shooter