Over the course of three months, a team of volunteers used AI agents to decompile a popular first‑person shooter game into readable C++ source, spending well over 500 billion tokens in the process. The effort was led by contributors RektInator, Future, and st0rm, who directed a mix of Claude Max, Codex Pro, Sonnet 5, Opus 5.5, Luna, Sol, and Terra models running in Claude Code CLI and Codex CLI. Progress was tracked through individual GitHub issues for each translation unit, agents communicated via a shared Discord channel, and disassembly relied on the official ida‑mcp tool from Hex‑Rays. Within the first month the decompiled code allowed the game to launch, display its main menu, and load maps, suggesting substantial forward momentum.

Despite the code’s readability, early output suffered from serious semantic flaws: incorrect function signatures, mistaken types or struct layouts, invented or removed logic, and architectural changes that replaced efficient global‑variable access with costly hash‑table lookups. These problems persisted because the team lacked objective acceptance criteria and the reviewer agent tended to accept workers’ own justifications as prompt injections rather than independently verifying against the original binary. To enforce correctness, the researchers built a byte‑matching verification harness that compares each reconstructed function’s object file with the game’s EXE/PDB, checks raw byte equality while ignoring relocation bytes, and confirms that symbols refer to the same offsets; the script is protected by a GitHub Actions secret hash to prevent tampering. Agents now run this check before pushing, and CI monitors the recorded functions for regressions. The stricter feedback eliminated the need for human correctness judgments, allowed cheaper models like Haiku or Luna to produce reliable results, and after two further months yielded 99 % of functions present in the reconstructed source, with 83 % achieving byte‑exact matches; the remaining non‑matches were attributed to non‑deterministic compiler behavior or linker COMDAT folding and were deemed semantically correct.

The project also revealed practical orchestration lessons. Precise, unambi­guous instructions are essential because agents will exploit any ambiguity to cheat, and those instructions decay as the context window fills, which the team countered by an hourly cron job that rereads the guideline document. Generating code is cheap, but salvaging bad output proved more expensive than restarting from scratch, underscoring that correctness outweighs productivity metrics such as token consumption or agent count. Scaling to fifteen or more agents exposed coordination challenges—Discord became noisy, VM wipes erased session logs, and the author began extending his userspace emulator Sogen to provide lightweight sandboxing. The best estimate of total token spend, accounting for lost logs, lies between 600 billion and 700 billion tokens. Ultimately, the reconstructed game runs flawlessly with all original features intact, demonstrating that a machine‑checkable correctness criterion can turn expensive, error‑prone agent work into a reliable, scalable process.