500B Tokens Later: Letting AI Agents Decompile a First-Person Shooter
davikr · hn · 2026-10-11
A detailed write-up of an experiment using AI agents to decompile a first-person shooter game binary, burning through roughly 500B tokens in the process. The author shows how agents autonomously parse disassembly, rebuild functions and data structures to recover game internals, and covers workflow design, context handling, and token costs — a real-world look at LLM agents tackling long-horizon reverse engineering.
More from coding & agent
- 10 GitHub repos that stop your AI agent from leaking keys and personal data — victor_explore · 2026-10-11
- Dev in prod: running a newsletter side project on a cheap persistent VM — davidcrawshaw · 2026-10-11
- AI assistant fixes GNOME dock broken by major update in about an hour — Moarkush · 2026-10-11
- GitHub CEO shares his Vibe Coding experiment building a game engine — zeeg · 2026-10-11
- DeepSeek Bot: open-source plugin turns DeepSeek Harness into persistent multi-bot teams — teortaxesTex · 2026-10-11
- 4K/60fps music visualizer coded and rendered entirely via ChatGPT — most_triumphant_yeah · 2026-10-11