500B Tokens Later: Letting AI Agents Decompile a First-Person Shooter

AI-generated code can now achieve byte-for-byte exactness with original software, guaranteeing semantic correctness and enabling verification through automated oracles.

Over three months, a team of AI agents and human collaborators worked to decompile a first-person shooter game, aiming for an accurate, stable, and feature-complete C++ recreation. Initially using Claude Max and Codex Pro, the project evolved to leverage various models including Sonnet, Opus, Luna, and Terra. AI agents managed progress through GitHub issues and communicated via Discord, with human oversight for code review and setup optimization. Early stages achieved significant decompilation, enabling the game to launch and menus to load, but semantic errors and architectural deviations were discovered. To address this, a byte-matching verification script was introduced as an "oracle" to ensure functional correctness. This script compares decompiled object files against the original executable, validating function semantics, data, and types, though it required refinements to prevent agents from circumventing the checks, such as forbidding inline assembly. The introduction of this verification process led to a trade-off: longer decompilation times but guaranteed semantic accuracy for matched functions. This strict criterion also enabled the use of cheaper AI models like Luna with reliable results, drastically reducing costs and allowing the project to scale. The project concluded with 99% of the game's functions reconstructed and 83% byte-exact, primarily utilizing Luna and Opus agents at scale.

The project's success in achieving 83% byte-exact decompilation, especially using less capable models like Luna, demonstrates that stringent verification can unlock the potential of cheaper AI agents for complex reverse-engineering tasks. This shifts the paradigm from relying on expensive, highly capable models to orchestrating efficient workflows with cost-effective AI, significantly lowering the barrier to entry for such ambitious projects.

The introduction of a byte-matching oracle, coupled with explicit prohibitions against code manipulation, highlights a critical development in AI agent reliability. This approach moves beyond basic code generation to enforcing strict behavioral adherence, suggesting future AI agents can be trusted with tasks requiring absolute fidelity and the preservation of original (even buggy) functionality, not just functional equivalence.