The Modern Warfare 2 AI agents used by a developer rebuilt the source code of the 2009 game, consuming more than 500 billion tokens over about three months. The project achieved 99% function coverage, with 83% matching the original executable byte-for-byte, and the game became fully playable.
What happened?
Maurice Heumann led the experiment using a fleet of AI agents based on Claude (Anthropic) and Codex (OpenAI) models. The agents analyzed the original 2009 game binary via tools like IDA Pro and Ghidra, through an ida-mcp server. Work started with four agents (three executors and one reviewer) and reached 17 agents working in parallel by the end of September.
Each translation unit received a GitHub issue, and all agents shared a Discord channel with CI notifications. Heumann estimated total consumption between 600 and 700 billion tokens, with costs ranging from about $1,200 (via subscriptions) to $200-250 thousand (via API).
Why does this matter?
The case of the Modern Warfare 2 AI agents demonstrates the progress of coding agents in complex reverse-engineering tasks. Matching decompilation (generating code that compiles byte-for-byte identical to the original) provides an objective success criterion that allows agents to iterate autonomously for hours. This accelerates projects that previously required human teams for months or years.
On the other hand, Activision intervened: videos shared by Heumann were disabled for copyright, and the site hosting the code went offline. The decompiled code remains private for legal reasons.
What changes in practice?
AI agents can already rebuild legacy software at scale, which could facilitate preservation of old games, vulnerability analysis, or ports. However, copyright issues remain the main limit. Similar projects, such as Rust rewrites of the same game (IW4L), also emerged with AI support, but without distributing original assets.
For developers and fans, the case shows both the potential and the risks of using AI in reverse engineering of commercial titles.
Image credit: Reproduction — Source: TechSpot / Maurice Heumann
By GeekikiBot
