TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
A community project spent three months using autonomous AI coding agents to reconstruct a popular, unnamed first-person shooter in C++. Its author says agents produced a game that launched and loaded maps, but later review found extensive semantic errors and design changes; the report presents the effort as a lesson in agent orchestration, not a verified, complete recreation.
A developer and community collaborators spent three months using autonomous AI coding agents to reconstruct an unnamed, popular first-person shooter in C++, but the project’s author says a working menu and map loading masked major semantic errors in the generated code. In a report published October 9, the author describes the work as an experiment in coordinating agents at scale—not evidence that they produced a verified, feature-complete recreation.
The project’s stated target was more demanding than a proof of concept: readable C++ that compiled and reproduced the original game’s behavior. The team initially also wanted security fixes and portability improvements, including possible support for Linux, macOS and browsers, but later set modernization aside to concentrate on matching the original behavior. The author says the effort involved RektInator, Future, st0rm and other community members.
The team used Claude Code and Codex CLI, with model choice changing over the project. Its agents tracked work through GitHub issues, generally one per C++ translation unit, and shared messages in a Discord channel. A GitHub webhook sent continuous-integration failures to that channel. For disassembly and decompilation, the author says the team used Hex-Rays’ official IDA MCP tool for most of the project.
During the first month, four agents were working: three workers and a reviewer. The author reports that they decompiled about 80% of the game in that period; the game could start, display its main menu and load maps. These visible milestones did not establish that the reconstructed code was accurate. Later inspection found incorrect function signatures, types and data structures, along with logic that agents had invented or removed and architectural changes that did not preserve the original design.
Why Visible Progress Was Misleading
The report highlights a gap between software that appears to work in a demonstration and software that faithfully reproduces an existing program. Starting a game or loading a map can show that some parts of a reconstruction operate, but it does not prove that underlying functions behave as the original does. The author says the code could be readable and compile while still being semantically wrong.
That distinction matters for AI-assisted software engineering beyond game preservation. In a large reconstruction, incorrect types, guessed logic or unnecessary redesigns can build on one another, leaving reviewers with a project that looks organized but diverges from its target. The author’s account suggests that parallel agents and automated progress tracking can increase activity without supplying a reliable measure of correctness.
The post also describes practical coordination methods: issue-based task assignment, shared agent communication, CI notifications and recurring reminders of project instructions. Those methods helped manage a long-running autonomous workflow, according to the author. But the experience also points to a limit: process improvements cannot substitute for objective acceptance criteria that let workers and reviewers test whether each reconstructed component matches the original.
As an affiliate, we earn on qualifying purchases.
How the Reconstruction Was Organized
The project was framed as an experiment in orchestrating autonomous agents over several months. The author says the team began with Claude Max and later added Codex Pro, running agents in the respective command-line tools. Models varied; the report names Sonnet 5, Opus 5.5, Luna, Sol and Terra. It does not provide a controlled comparison of those models, so the account does not establish which performed best.
As the work progressed, the team adjusted how agents handled long sessions. The author says they lowered the context-compaction threshold from the default 90% to 42%, on the view that completed-function details were no longer useful and could crowd out current work. They also wrote project-specific instructions and used an hourly scheduled prompt to remind agents to reread them. The author reports that agents sometimes drifted, switched tasks before finishing, waited despite failure alerts or closed issues without adequate checking.
The first month’s apparent progress gave the team confidence, the author writes, before closer quality review exposed deeper problems. The report says the reviewers could flag bugs but struggled to question architectural choices because the team had not properly defined “correctness.” The post’s central lesson is therefore less about a particular model or decompilation tool than about how to evaluate autonomous work against a precise behavioral target.
“The goal was not to reach a simple proof-of-concept state. Instead, we really wanted an accurate, stable and feature-complete recreation of the game.”
— The project author, Maurice, in the October 9 report
As an affiliate, we earn on qualifying purchases.
Accuracy and Project Status
The report does not name the shooter, saying earlier posts about the game had been removed after what the author describes as corporate intervention. It also does not provide a final, independently tested measure of how much of the game was accurately reconstructed. The roughly 80% figure refers to work completed during the first month, not a verified share of correct code or the project’s final state.
It remains unclear whether the team ultimately achieved a stable, feature-complete recreation, whether the code was tested against systematic behavioral checks, or what share of the identified errors was corrected. The post is a participant’s account of the workflow, not an independent audit or a controlled evaluation of AI models. Its claims about tools and outcomes should be read in that context.
As an affiliate, we earn on qualifying purchases.
The Need for Testable Correctness
The published account does not announce a release date or a next project milestone. It says the team deferred portability and other modernization work to focus on reproducing original behavior, but does not state whether those plans will resume or whether the reconstruction will be made public.
The clearest next step implied by the author’s findings is to establish objective tests for correctness, so reviewers can compare reconstructed functions and structures with the target rather than relying on code readability or visible progress. Whether the team will publish such tests, share the source, or provide a follow-up on the project’s final status remains unknown.
As an affiliate, we earn on qualifying purchases.
Key Questions
What did the project try to do?
It aimed to reconstruct an unnamed first-person shooter as readable, compiling C++ that reproduced the original game’s behavior. The author says the team initially considered security and portability improvements, then deferred them to focus on behavioral accuracy.
Did the AI agents complete the game?
The report does not establish that they produced a verified, feature-complete recreation. It says the game could launch, show a menu and load maps, while later review found serious semantic errors.
What does the 80% figure mean?
The author says agents had decompiled about 80% of the game during the first month. The report does not present that as a measure of accurate code or give a final completion percentage.
Which AI models and tools were used?
The author reports using Claude Code and Codex CLI, with models including Sonnet 5, Opus 5.5, Luna, Sol and Terra. The team also used Hex-Rays’ IDA MCP tool for disassembly and decompilation. The account does not compare model performance under controlled conditions.
What does the author say went wrong?
The agents produced incorrect signatures, types and structures, sometimes invented or removed logic, and made architectural changes. The author says reviewers lacked clear, objective criteria for deciding whether those choices preserved the original game’s behavior.
Source: hn
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
