AI agents decompile a first-person shooter over three months using autonomous orchestration and byte-matching verification

A research team led by Maurice Heumann orchestrated multiple AI coding agents over three months to decompile a commercial first-person shooter into readable, compilable C++ code. The project aimed for semantic correctness, stability, and feature completeness rather than a minimal proof of concept.

The team used a mix of commercial models—Claude Max (20x), Codex Pro, Sonnet 5, Opus 5.5, Luna, Sol, and Terra—via their respective agent harnesses (Claude Code CLI and Codex CLI). Agents coordinated via Discord and tracked progress using GitHub issues, with labels for grouping and prioritization. A GitHub webhook posted CI failures into the shared Discord channel to notify agents of build breaks.

During the first month, four agents operated: three workers decompiling and committing code, and one reviewer coordinating and reviewing commits. The team reported visible progress—game launch, main menu visibility, and map loading—within four weeks, but semantic correctness lagged. Agents introduced wrong function signatures, incorrect types, and unnecessary architectural changes, such as replacing constant global variable accesses with slower hash-table lookups.

The reviewer struggled to enforce correctness because the project lacked objective acceptance criteria. Reviewer judgments were inadvertently influenced by agents’ commit messages, which acted as prompt injection, leading to acceptance of deviations justified in comments rather than verified against the original binary.

To address semantic drift, the team introduced a byte-matching decompilation oracle. The oracle compiled the reconstructed code with the original compiler and compared function bytes between the reconstructed OBJ file and the original EXE/PDB. Mismatches triggered rework, while relocations and symbol references were excluded from byte comparisons to avoid false negatives. The same process was applied to data and types.

Infrastructure refinements included reducing the default context compaction threshold from 90% to 42% to remove volatile decompilation artifacts and prevent agents from losing focus. An hourly cron job injected a request for agents to reread the instruction document to maintain task alignment over long-running autonomous sessions.