500B tokens later: Letting AI agents decompile a first-person shooter
Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated.
A months-long autonomous agent swarm (3 workers + 1 reviewer on Claude Code and Codex CLI) burned 500B tokens to decompile ~80% of a shooter into compiling C++, coordinating via GitHub issues (one per translation unit) and a shared Discord channel for agent-to-agent and human-in-the-loop messaging. The counterintuitive production lesson: they slashed the context compaction threshold from 90% to 42% because stale decompiled functions poison the context, so for long-horizon agent work aggressive early compaction cuts token cost and keeps agents focused rather than letting context fill naturally.
A small autonomous agent setup consumed roughly 500B tokens over three months decompiling an unnamed popular FPS into readable C++, reaching about 80% completion with the game launching, showing the main menu, and loading maps. The useful lesson is operational: long-running agents worked best with GitHub issues per translation unit, Discord/CI feedback loops, Hex-Rays ida-mcp, a reviewer agent, and aggressive context compaction to discard stale reverse-engineering state.
AI vs. AI Debate
“The summary is mostly accurate but over-indexes on the 42% compaction tweak while omitting key operational details like ida-mcp, CI feedback, and the concrete validation that the reconstructed game could launch and load maps.”
“While I compressed the toolchain and validation details for brevity, the compaction insight is explicitly framed as the article's "counterintuitive production lesson," so my emphasis correctly tracks the author's own highlighted takeaway rather than over-indexing.”