AI news · October 13, 2026
A 500B-token game decompilation project reaches 83% byte-exact code
The project reports 99% of functions present and 83% byte exact after nearly 3 months, using 14 Luna and 2 Opus agents; the useful idea is the acceptance harness, not the token bill.
- functions reported byte exact
- 83%
- functions present in source
- 99%
- agents named in final weeks
- 16
Maurice Heumann described a three-month attempt to recreate a first-person shooter with coding agents. The team began with four workers and a reviewer, reached a visible game menu and maps, then found that readable code was still semantically wrong. It replaced broad human review with a compiler and comparison script that checked reconstructed object files against the original binary, while a hash protected the verification script from agent edits.
The final report says 99% of functions were present and 83% were byte exact, with 14 Luna and two Opus 5.5 agents used in the final weeks. The author later estimates total use at 600B to 700B tokens, so the 500B title is only a rough marker. The useful lesson for normal software is smaller: define correctness before scaling agents, make the check automatic and reject work that cannot pass it. Progress dashboards alone hid serious semantic errors.
What you can do with it
Borrow the harness idea on a small module. Define an observable contract, create tests that compare the old and new implementation, and let an agent work only on tasks with a machine-checkable result. Keep workers on separate branches, record cost per accepted change and stop if the evaluator cannot distinguish a good result from a plausible one.
Our take
This is a strong case study in why more agents do not fix a vague definition of done. The final percentages are the author's report and the code remains private, but the failure pattern is familiar: readable output can still be wrong. A narrow evaluator can make cheaper models useful, while an unmeasured swarm only creates more review work.
Links Project reportHN discussion
Source: Maurice Heumann / Hacker News — Made With Models writes the brief; the reporting is theirs.