- Published on
500B Tokens Later: Letting AI Agents Decompile a First-Person Shooter
- Authors
- Name
- DeepHorus
Over three months, a team of developers and autonomous AI agents decompiled a popular first-person shooter into C++. The effort consumed an estimated 500 billion tokens and involved RektInator, Future, st0rm, and other community members. The goal was not a proof-of-concept but an accurate, stable, feature-complete recreation. The project also served as a large-scale experiment in orchestrating AI agents over months.
The game's identity was not disclosed; earlier blog posts were removed after what the author described as corporate America ruining the fun. The article focuses on AI orchestration, infrastructure, setup, and harness optimization rather than decompilation process. The team wanted readable C++ source that compiles, plus security, bug fixes, portability. They later deferred modernization and portability to reconstruct original behavior.
The team began with Claude Max 20x, then added Codex Pro, running both subscriptions simultaneously. Model choice varied: Sonnet 5 most often, with Opus 5.5, Luna, Sol, Terra also used. Claude agents ran in Claude Code CLI, Codex agents in Codex CLI. Other harnesses barely mattered, so defaults stuck.
Using GitHub CLI, agents managed issues to track progress, one issue per translation unit .cpp file. Labels grouped and prioritized issues. Agents communicated via Discord in one shared channel, enabling agent-to-agent and human-to-agent communication. A GitHub webhook posted CI failures into the channel so agents knew when something broke.
Agents used official ida-mcp by Hex-Rays almost the entire time. It was described as stable, headless, and supportive of everything needed. First month: four agentsâthree workers decompiling and committing, one reviewer passively coordinating and reviewing commits to flag bugs. They decompiled about 80% of the game; the game launched, main menu visible, maps loaded.
The team spent weeks optimizing setup. They reduced token consumption by triggering earlier compactions, lowering threshold from default 90% context fill to 42%. Decompilation creates volatile information: once a function is decompiled, it becomes irrelevant and can be removed. Earlier compactions help remove junk and mitigate focus loss. Agents drift even within one compaction cycle as context fills.
Agents sometimes moved to another function before finishing the previous one. They idled while watching CI despite receiving Discord failure notifications. Occasionally, they closed issues without thoroughly checking whether work was complete. Working side by side in a terminal can steer agents to prevent this, but when letting them run autonomously, such issues persist. The project remains a lesson in orchestration, monitoring, and context management.
The team's next steps likely involve refining compaction thresholds, issue tracking, and communication protocols. The experience suggests AI agents can achieve substantial decompilation progress, but human oversight remains critical. As models and harnesses improve, similar projects may target feature-complete recreations with less drift. For now, the 500-billion-token effort shows both the promise and limits of autonomous agent swarms.
