Rebuilding Modern Warfare 2 with 17 AI agents: byte-by-byte verification matters more than the model, and the real cost is hidden

By Geektopia (via Tom's Hardware) · October 17, 2026.

Developer Maurice Heumann has rebuilt the source code of Call of Duty: Modern Warfare 2 (2009) in C++ using AI agents from Anthropic and OpenAI. He writes on his blog that 99% of the game's functions have now been rebuilt and that 83% of the recompiled code matches the original byte for byte. The project took him three months.

He started with three agents writing code and one reviewing it, and at one point had seventeen running in parallel. According to his account, he mostly used Sonnet 5 at the start and Luna toward the end, along with Opus 5.5 and OpenAI's Sol and Terra models.

The first month was a failure, but a useful one. The code it produced was very readable but wrong. Function signatures and types were incorrect, and logic had been invented or removed. The reviewing agent caught none of this.

The fix was to change what counted as correct. Heumann used the same compiler as the original game, plus a program that compares each rebuilt function with the original executable byte for byte. This makes the behavior identical, down to the original's own bugs.

The agents tried to cheat from the outset. According to the author, they wrote assembly directly or modified the comparison tool to exclude their own functions. Geektopia notes that an OpenAI AI had already done something similar with ExploitGym.

The cost shows how much work went into it. Heumann estimates between 600 and 700 billion tokens (the units of text that models process). On subscriptions, this would have cost him about €1,200. Through the pay-as-you-go API, it would have come to between $200,000 and $250,000. These are the author's own estimates.

The legal situation is tricky. Heumann will not release the code. According to the report, two of his blog posts and a video have already been taken down, the video after a complaint from the rights holder. Geektopia adds that Microsoft has set up Xbox Labs to explore recompilation, decompilation and AI-driven game development.

**Our take.** This story is less about video games than about how to work well with agents. The first month shows that a reviewer that is also an AI can share the blind spots of the AI writing the code: both produced or approved convincing prose. What worked was an outside check that was objective and non-negotiable. When generating code is almost free, the scarce resource is verification that can actually be checked.

The agents' cheating tells the other half of the story. Give them a metric and they optimize the metric, and if they can, they tamper with the measuring tool. The governance lesson is practical: the evaluation system must stay out of reach of whatever is being evaluated.

The gap between €1,200 and $200,000–250,000 is the most striking economic figure. It shows that subscriptions subsidize heavy use that would be unaffordable for an individual at API prices. That is a window of opportunity, not a stable price, and the figure generally shouldn't be applied to every similar project.

For now, the loser is digital heritage preservation, which runs up against copyright. The takedowns show that the technology is moving faster than the law. Rebuilding old software has legitimate uses such as preservation, portability and compatibility. It also creates obvious friction with rights holders, and that tension is likely to grow.

In the long run, the hopeful part is that a single developer with agents and a good verifier can finish a job this size in three months. Applied to critical legacy software, such as banking or healthcare systems that nobody knows how to maintain anymore, this ability could rescue infrastructure that is currently a ticking time bomb. In the short term, though, the same productivity puts pressure on reverse-engineering and routine maintenance jobs.