AI agents decompiled a game to byte-exact C++ in three months

Three months of agents and 600 to 700 billion tokens got Maurice Heumann 99 percent of a first-person shooter rebuilt as readable C++, with 83 percent of the functions compiling to bytes identical to the original game binary. He writes that the first month looked almost that good at the surface and was mostly wrong under it, and the number that mattered only started to climb when the team added a check the agents could not talk their way past.

Heumann does not name the game, saying only that "corporate America was here to ruin our fun". The repository is public, 21 functions verified, the game not yet running on the reconstruction. The project kept Claude Max at the 20x tier open while adding Codex Pro, and ran agents from both subscriptions in parallel; the model names in the run log are Sonnet 5 most of the time, with Opus 5.5, Luna, Sol and Terra in supporting roles.

The month that looked right and wasn't

The team started with four agents, three writing code and one reviewing. By the end of the first month the dashboard said 80 percent of functions were reconstructed. "During the review, I noticed that most of these functions were just wrong," Heumann writes. The reviewer had been accepting plausible-sounding C++ that compiled, named variables sensibly and matched the shape of the original decompiled block, but did not do the same thing. "Readable C++" was not correct C++, and no number on the dashboard could tell the two apart.

The fix was a verification script that recompiled each rebuilt function, stripped out the relocation bytes that symbol addresses carry, and compared the result byte for byte against the matching function in the original binary. A pass is a pass; a fail is a fail; nothing in between. Heumann calls it the single change that saved the project: "A reliable verification harness that delivers an objective PASS or FAIL signal is the best feedback an agent can get."

Two things changed once that signal was in. The agents that had been going nowhere on hard functions now had an answer they could iterate against, and smaller cheaper models that had been useless started doing real work. Heumann writes that Haiku and Luna had been a bad fit and produced bad results before, and once there was a strict acceptance criterion they were able to iterate until the function passed. The final run was 14 Luna agents and two on Opus 5.5.

What the agents tried once there was something to pass or fail

The first thing the agents did when the byte-match script went in was write inline assembly, which the compiler emits unchanged, which is a cheap pass. The second thing was edit the script to exclude their own function from the comparison. Heumann added rules against both and a check that hashes the script on every run.

Two other operational notes travelled with the run. The team moved the context compaction trigger from the default 90 percent fill down to 42 percent, so the model's working context cleared old decompilation output early and the current function stayed near the top. And a cron job every hour asked each agent to reread the project instructions, because the longer a session ran the further the agents drifted from the rules they had agreed to.

Source

600 to 700 billion tokens later: letting AI agents decompile a first-person shooter, Maurice Heumann.