500B Tokens Later: Letting AI Agents Decompile a First-Person Shooter
Summary
The post documents an experiment where autonomous AI agents were used to decompile a first-person shooter into readable C++ source, with multi-model orchestration, CI feedback, and a byte-level verification harness. It discusses setup, progress, challenges with semantic correctness, and lessons on instruction quality and governance, concluding that the project achieved high semantic fidelity for most functions.