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

During the last three months, the reporter spent some time and tokens decompiling a popular first-person shooter. The goal was not just to achieve a simple proof-of-concept state, but rather an accurate, stable, and feature-complete recreation of the game. The reporter had previously written two posts about this project, which have since been removed. The game in question was made popular by corporate America, but that's irrelevant to this post.

The reporter aimed for an accurate decompilation of the game to C++, focusing on semantic correctness, readable code, and portability improvements. The overall goal was to learn how to effectively orchestrate autonomous AI agents over several months.

The initial setup involved using Claude Max (20x) and Codex Pro, running both subscriptions simultaneously. Model choice varied, with Sonnet 5, Opus 5.5, Luna, Sol, and Terra being used frequently. Claude agents ran in Claude Code CLI, while Codex agents used Codex CLI. Other agent harnesses were also tried, but the default settings worked fine.

Progress tracking was managed using GitHub CLI, with one issue per translation unit (.cpp file) and labels to group and prioritize issues. Agents communicated via Discord, allowing agent-to-agent and human-to-agent communication. GitHub webhooks posted CI failures into the shared channel, notifying agents of any issues.

The first month saw four agents running, with three workers decompiling and committing code, and one reviewer agent coordinating and reviewing commits. By the end of the month, the agents had decompiled about 80% of the game, with visible progress including launching the game, viewing the main menu, and loading maps. The team then spent time optimizing their setup, reducing token consumption by triggering earlier compactions. They also noticed agents tended to lose focus over time, drifting from the task at hand.

Despite the progress, the reporter's initial optimism about the quality of the decompilation proved misguided. The code was readable but semantically incorrect, with agents making errors in function signatures, types, and struct layouts. They also introduced unnecessary architectural changes, such as turning constant memory access into more expensive hash tables.

The reviewer agent helped catch some bugs but was not effective in evaluating architectural decisions, as objective acceptance criteria were not established. This lack of clear criteria made it difficult to judge whether changes were correct or incorrect.