Post
Taulet's avatar
⋮

500+ Billion Tokens Later: Letting AI Agents Decompile a First-Person Shooter

A months-long decompilation project became a report on engineering the agent workflow around it. The team used Claude Code and Codex CLIs, tracked work by translation unit in GitHub issues, coordinated agents and people in a shared Discord channel, and routed CI failures into that channel. They used IDA’s headless MCP integration for disassembly, and ultimately deferred modernization and portability to focus on reproducing the original game’s behavior.

The 43-comment HN discussion questioned how to define a faithful decompilation: one commenter proposed accepting semantically equivalent code with tests run against the original, while others warned that test coverage can miss behavior such as overflow. Another suggested mechanical decompilation plus SMT equivalence checks before asking an agent to clean up the code. The author’s post is useful as a concrete account of agent orchestration, not evidence that a particular model or harness guarantees correctness.