Maurice Heumann says he and community collaborators spent three months using AI agents to decompile a commercial first-person shooter into C++. He estimates the project consumed more than 500 billion tokens, but that figure has not been independently audited. The current post does not name the game; preserved earlier wording identifies it as Call of Duty: Modern Warfare 2 (2009).
What the project set out to do
In an October 9, 2026 post, project author Maurice Heumann described a sustained effort to reconstruct a commercial shooter as readable, compiling C++ while accurately reproducing the original game’s behavior. Modernization and portability were later deferred. The work was also an experiment in coordinating autonomous agents over months, rather than a one-prompt game-generation demonstration. Heumann summed up the technical aim: “We aimed at an accurate decompilation of the game to C++.” Heumann’s project account
As an Amazon Associate I earn from qualifying purchases.
He reports using Claude Max and Codex Pro subscriptions, with agents running in Claude Code CLI and Codex CLI. GitHub issues organized work around translation units, agents coordinated in a shared Discord channel, and Hex-Rays’ ida-mcp supported disassembly and decompilation. Heumann also described learning how to coordinate agents over a long project as a central goal.
How many tokens did it take?
Heumann’s rounded headline estimate is more than 500 billion tokens for the project. It is an author-reported figure, not a result independently audited in the available account or secondary coverage; no independent token telemetry is supplied. It should not be read as a dollar-cost estimate: model use and subscription arrangements vary, and the reported total does not establish what the project cost.
#1 Best Overall
Was the game really Modern Warfare 2?
The October 9 post does not name the shooter. Preserved earlier wording and community-linked archive references identify it as Call of Duty: Modern Warfare 2 (2009), but that identification is not stated in the current post. The attribution should therefore be treated as coming from earlier wording, not as a detail confirmed in Heumann’s current account. Contemporaneous discussion preserving the earlier identification
How did the team check whether the reconstructed code was right?
Heumann describes using a harness that could compare reconstructed behavior against the original in a machine-checkable way. That matters because compiling successfully, or appearing to run correctly, does not prove that the internals reproduce the original. A verification harness makes differences detectable under the comparisons it performs; it does not, by itself, establish equivalence for every possible input or behavior.
Byte matching versus functional equivalence
Discussion around the project highlights a consequential choice: require byte-identical output, or accept functionally equivalent results. Byte equality can expose divergences that a finite set of behavior tests may miss, but it imposes a stricter target and demands close attention to exactness, including compiler behavior. Functional equivalence can allow implementation differences while requiring a clear account of which behaviors the harness tests. Neither label alone tells readers how comprehensive or independent the checks are.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errorsWhat to evaluate in a verification claim
- Comparison target: Does the harness require byte-for-byte matching or test functional behavior?
- Coverage: Which inputs and execution paths are checked, and what remains outside the tests?
- Independence: Is the comparison grounded in the original program, or only in expectations derived from the reconstructed code?
- Project trade-offs: Is the priority fidelity to original behavior, or readable, portable, modernized code? Heumann says modernization and portability were deferred in this effort.
- Verified progress: How much of the claimed output has actually passed checks, rather than merely been generated by agents?
What the project demonstrates—and what it does not
The reported effort illustrates that long-running agent work depends on more than model output: task tracking, coordination, reverse-engineering tools and a verification harness all played a role. Its scale is notable as a project account, but token volume measures reported consumption, not correctness or independently verified progress. The available sources also do not establish a definitive legal ruling about this particular project; no legal conclusion follows from the technical description.
Quick Recap
Best Value
Rank #4
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
