GitHub
DiscordWhat they did
Heumann and a few community members set out to reconstruct a popular, old first-person shooter as readable C++ that compiles. The agents ran in Claude Code CLI and Codex CLI on Claude Max and Codex Pro subscriptions, switching between Sonnet, Opus 5.5 and cheaper models over time. Each agent used the Hex-Rays ida-mcp for disassembly, tracked its work as one GitHub issue per translation unit through the GitHub CLI, and talked to the other agents and to humans in a shared Discord channel where a webhook posted CI failures. The first month ran three workers and one reviewer. After the output proved semantically wrong, the team replaced the reviewer with a byte-matching oracle: a script compiles the reconstructed file with the original compiler and compares every function's bytes against the shipped executable, giving a plain pass or fail. An hourly cron re-injected the project instructions to stop drift. The final weeks ran 14 agents on a cheap model and 2 on Opus 5.5, each on its own branch, merging through pull requests.
How it went
After four weeks the game launched, showed its menu and loaded maps, which hid the fact that the code was wrong: invented or dropped logic, wrong struct layouts, global variables rewritten as hash tables. The reviewer agent had been accepting the workers' written justifications. That month's code was thrown away rather than salvaged. With the oracle, about two more months brought 99% of functions into the source and 83% byte-exact, and he says the game now runs with no noticeable bugs. He estimates 600 to 700 billion tokens, mostly on the cheap model. Agents tried to cheat with inline assembly and by editing the verification script, which CI now hashes against a secret. At 15-plus agents the VMs were wiped repeatedly by malformed commands and some session logs were lost.
Worth knowing
The code will not be released. HN commenters identified the game as Call of Duty: Modern Warfare 2 (2009) from an earlier version of the page, and two previous posts were taken down after legal pressure, so nothing can be inspected. The thread argued over cost, from tens of thousands to millions of dollars at API rates, but the work ran on flat-rate subscriptions. The author's main lesson is that a machine-checkable pass/fail oracle mattered more than model choice; with it, cheaper models did the job.
