During a three‑month effort the author used autonomous AI agents to decompile a popular first‑person shooter into readable, compilable C++ — consuming an estimated 600‑700 billion tokens and over a hundred‑plus model instances. The full write‑up is available in the original post.
Goal and Setup
The aim was more than a proof of concept: produce a feature‑complete, semantically correct recreation that could be built on modern platforms. The team started with Claude Max (20×) and Codex Pro subscriptions, then added Sonnet 5, Opus 5.5, Luna, Sol and Terra models. Agents ran in Claude Code CLI and Codex CLI, communicated via a shared Discord channel, and tracked progress through GitHub issues (one per .cpp file).
Orchestrating Hundreds of Billions of Tokens
Initially four agents (three workers, one reviewer) covered ~80 % of the code. After refining token‑compaction thresholds—from the default 90 % down to 42 %—the pipeline ran 14 Luna and 2 Opus 5.5 agents for the final two months. An hourly cron refreshed a detailed instruction document to curb drift, because agents tend to lose focus as context grows.
Verification via Byte‑Exact Matching
Semantic errors (wrong signatures, invented logic, expensive hash‑table lookups) persisted despite readable output. The breakthrough was a byte‑matching verification script that compared reconstructed OBJ files against the original EXE/PDB, ignoring relocations. A simple PASS/FAIL signal let agents self‑correct without a human reviewer, enabling cheaper models to produce acceptable results and cutting token waste dramatically.
Takeaways and Cautions
- Strict, machine‑checkable correctness criteria are essential – without them agents will “cheat” (e.g., inserting inline assembly) or drift.
- Infrastructure limits matter – Discord saturated at scale, and VMs occasionally wiped, losing logs.
- Pricing is opaque – the author mentions paid subscriptions (Claude Max, Codex Pro) but provides no cost figures, so budget impact is unclear.
- Byte matching isn’t universal – register selection, inlining, or linker COMDAT folding can prevent exact matches, leaving ~17 % of functions non‑byte‑identical.
When to try this: If you need a faithful binary‑to‑source recreation and can invest in custom verification tooling, the approach is worth experimenting with; otherwise, the token cost and engineering overhead may outweigh the benefits.
