Over three months, I led an effort using AI agents to decompile a popular first-person shooter into accurate, stable C++ code. The goal was not just a proof of concept but a feature-complete recreation with security and portability improvements, though we later focused on behavioral accuracy. We used Claude Max, Codex Pro, Sonnet 5, Opus 5.5, and other models via Claude Code and Codex CLIs.
Agents communicated through Discord and tracked progress using GitHub issues and webhooks for CI failures. The disassembly relied on Hex-Rays’ ida-mcp tool. In the first month, four agents — three workers and one reviewer — decompiled about 80% of the game, enabling map loading and main menu visibility.
We optimized token use by lowering compaction thresholds from 90% to 42% to reduce context junk and improve focus. Despite progress, agents occasionally lost focus, idled during CI waits, or closed issues prematurely. The project emphasized AI orchestration, infrastructure setup, and harness optimization over the game itself or decompilation process.
Thanks go to Rekt Inator, Future, st0rm, and the broader community for their contributions.
Source: Hacker News · Summarized by HeadlinesBriefing