200 Billion Tokens Later: AI Agents Decompile Modern Warfare 2

Summary

A month-long experiment used Claude Code with four cooperating agents to decompile Call of Duty: Modern Warfare 2 (2009) into C++ as close to the original as possible, with planned portability, security, and stability improvements. Three worker agents handled separate game subsystems, while an overseer reviewed commits; they coordinated through Discord, GitHub issues, webhooks, CI, Ghidra, and IDA Pro MCP servers. After about four weeks, the private project had produced nearly 7,000 commits and decompiled roughly 5,588 of 16,324 functions, or 34%, and the game could launch. Map loading still failed because important subsystems remained incomplete or unwired. The agents had used 199.8 billion tokens. GitHub issues worked better than a large status file, and Discord enabled human interaction, but the agents repeatedly produced overly long messages, ran unnecessary local test suites, avoided larger blockers, and sometimes stopped because they believed they were running out of context. The author found that assigning mechanical decompilation, modernization, and portability together caused agents to lose priorities after repeated compaction; separating those phases worked better. Switching from Opus 5 to Sonnet 5 produced no noticeable loss for this task, while PostCompact hooks helped preserve rules. The experiment was considered successful overall, but long multi-week sessions exposed memory, focus, and coordination problems beyond those seen in shorter sessions.