Autonomous AI agents decompile full game using 600B tokens

Over three months, a developer orchestrated multi-agent AI workflows to decompile a first-person shooter game into C++. The team initially combined Claude Max (20x) and Codex Pro subscriptions using Claude Code CLI and Codex CLI. They utilized Sonnet 5, Opus 5.5, Luna, Sol, and Terra models alongside the official Hex-Rays ida-mcp tool to process functions tracked as GitHub issues.

To manage token overhead and context drift across multi-month execution, the team reduced the context compaction threshold from 90% down to 42%. They also ran an hourly cron job that forced agents to re-read instruction documentation. The initial architecture used three worker agents and one reviewer agent coordinated via Discord. However, after four weeks of work, the output was semantically flawed—replacing global variable lookups with slow hash tables—because worker justifications acted as indirect prompt injections on the reviewer.

To enforce exact standard behavior, the team abandoned agent reviewers for a deterministic CI pipeline using the original game compiler, binary EXE, and PDB files to byte-match compiled OBJ functions. Verbal rules banned inline assembly, naked functions, and code-patching, while the test script was secured with a hashed GitHub Actions secret. This strict automated harness allowed less capable models like Haiku and Luna to reliably produce correct output.

During two subsequent months of automated execution, a fleet scaling up past 15 agents—primarily 14 Luna and 2 Opus 5.5 agents—brought 99% of functions into reconstructed source files, achieving 83% byte-exact matching and a fully playable game. To resolve agent command errors that wiped virtual machines, the researcher began extending the Sogen emulator for sandbox isolation. Total consumption reached an estimated 600 to 700 billion tokens, including about 300 billion Luna tokens (equivalent to 3 billion Astra tokens at a 100:1 ratio).