Microsoft is creating Xbox Labs, a small group inside Xbox that will explore the fast-growing practice of decompiling and recompiling games, much of it done with AI help. The news comes from an internal-style note from Xbox CEO Asha Sharma, reported by The Verge, which describes Microsoft as "officially embracing the chaos of this new AI-powered era of game modding." This post covers what was said, how the technique works, why the modding scene is split, and the questions Microsoft has not answered.
If you follow AI coding, the interesting line is not about games at all. Sharma argues games are a natural next workload for agents.
TL;DR: what Xbox announced
| Question | Answer |
|---|---|
| What is it? | Xbox Labs, a small group of game makers, engineers and scientists |
| Mandate | Turn emerging ideas (decomps, recomps, ports, remixes) into things players and developers can experience |
| Approach | "Build in the open," learn from what people do, measure success by whether more people play and create |
| Who runs it? | Per the note as quoted by The Verge, a leader named Scott was asked to stand it up |
| Tools | Xbox says it will share the tools it builds with developers across the industry |
| Open question | How Microsoft responds when modders redistribute its IP; the company did not respond to The Verge |
| Unchanged | Xbox says its priorities are not changing and that it remains committed to making great games, its hardware and responsible stewardship of IP |
What did Asha Sharma actually say?
The Verge quotes Sharma saying Xbox Labs will build in the open and measure success "by whether more people are playing and creating." She points to experiments in recent weeks where Windows PC builds were translated to run on Apple Silicon, or ported to run in a browser through WebAssembly and WebGPU, the latter now approaching 90% global browser support. She adds that many of the biggest games in the world "have been remixed into the source code of another."
These experiments, she writes, "challenge long-held assumptions about our industry: the tether between software and hardware, the boundary between creation and play, and how people discover and experience games." She calls the outcomes "unmistakably good outcomes for players, even when (and sometimes because) they challenge our existing business models."
The agent argument is the most relevant for readers of this blog: "Coding became the first major workload for agents because they can write, run, and test software. Games share many of those same properties," and people spend more than 72 billion hours a year playing Xbox games, "roughly five times the estimated time spent writing code worldwide."
The Verge's Tom Warren adds that modders are now doing what Microsoft has done for decades with Xbox backward compatibility, "but they're doing it without the business model restrictions that Microsoft is tied to," and writes that it feels like "we're just a model or two away from even more breakthroughs."
How does game recompilation work?
Two terms get mixed up. Decompilation rebuilds human-readable source from a compiled binary, which takes years of manual effort for a large game because someone must work out what every function does. Static recompilation instead translates a game's machine code from one processor's instruction set to another ahead of time. As PC Gamer explains, unlike emulation, it does not simulate the original hardware as a whole.
The best-known hobby tool is N64 Recomp, announced a little over two years ago, which has been used for PC ports of Banjo-Kazooie, Star Fox 64 and Mario Kart 64, per PC Gamer. Similar projects exist for other consoles, including the Xbox 360. PC Gamer also notes that the freshly written code in these projects does not contain original assets; players provide their own ROM or game files, which is why legal exposure is debated differently from piracy.
AI changes the economics. A model that can read disassembly and propose function names and structure shortcuts the slowest part of the job. That is exactly the workload Sharma describes: write, run, test, repeat.
Why is the modding community split?
PC Gamer describes the AI question as an "ethical, rather than legal, battleground" in the last few months as "vibe coded" projects appear almost daily. Its example is Donkey Kong 64: a second project with the name DK64 Recompiled was announced in late June by members of the DK64 randomizer community, who said they would take ownership because another concurrent recompilation was "heavily leaning on AI vibe code" and would be harder to manage over time. In some cases, PC Gamer reports, ports written with tools like Claude are ostracized or ignored.
So Xbox is embracing a technique that parts of the community it is courting actively resist. That tension, speed and reach versus craftsmanship and maintainability, mirrors debates we have covered in open-source software, where projects like QGIS banned AI-assisted contributions and COSMIC desktop banned AI-generated code.
What are the IP and business questions?
The Verge says the bigger question is how Microsoft will respond to decompiles and recompiles where modders take its IP and make it freely available, and that it asked Microsoft to comment on unauthorized IP use without a reply in time. Sharma's own note keeps one foot in each camp: it welcomes the experiments and also reaffirms "being responsible stewards of IP and the communities built around it."
Reading the note, we see three possible directions, which are our inference rather than anything Microsoft has said:
- Official tooling. Xbox ships its own AI-assisted porting and backward-compatibility tools for its catalog.
- Licensed community projects. Rights holders bless certain recomps, as some publishers have done with fan efforts.
- Enforcement. Takedowns for ports of modern console titles that appear where they should not.
Which one wins depends on what Xbox Labs builds, and Sharma only promised to share tools and learnings.
What does this mean for AI builders?
- Games are being pitched as an agent testbed. Deterministic, testable, huge in scale: an obvious benchmark domain. We noted a related push at Epic Games, which ordered an AI-first build of its Disney shooter.
- Reverse engineering is an AI sweet spot, and a contested one. Our piece on the PS5 jailbreak and whether AI really cracked it shows how hard it is to separate hype from verified results in this area.
- Platform vendors are re-evaluating the hardware tether. If a game can run anywhere, console exclusives and storefront lock-in weaken, which relates to Xbox's wider restructuring and Game Pass changes.
- Agent-written code needs verification. A port that "runs" is not a port that is correct, a theme we covered in the verification tax on AI output.
What are people asking?
Does this mean Xbox games are coming to Mac and the browser? Sharma cites Windows builds translated to Apple Silicon and ports running in a browser as examples of what modders have done, not as products Microsoft is shipping. No release is announced.
Can I use AI to port a game myself? Technically the tools exist, and PC Gamer shows hobbyists doing it. Practically, expect scrutiny: some community servers reportedly decline to promote recomps where AI wrote a significant share, and publishers have their own views on redistribution. Keep any project asset-free, as the recomp scene does, and do not distribute game files.
Is AI really good at reverse engineering? It speeds up naming and structuring code, but results need testing against the original behavior. Treat claims of one-shot ports with skepticism until someone plays the game for hours.
Why would a platform owner want this? Backward compatibility has been a pillar of Xbox for years. If AI cuts the cost of making old games run on new hardware, the economics of preservation change, and a platform owner that owns the catalog is best placed to benefit. That reading is ours, not Microsoft's.
A practical test for builders
If you build agents, the recomp scene is a useful case study because success is easy to define: does the ported game behave like the original? That makes it a natural place to apply the write, run and test loop Sharma describes. A sensible evaluation would give an agent a small, legally clear target (an open-source or homebrew game), require a reproducible build, and score it on frame-by-frame behavior against the original. The same discipline applies to any migration job, from porting a legacy service to a new language to translating a build system: define equivalence first, then let the agent iterate against it.
It also shows the limits. Unit tests rarely exist for decades-old games, so equivalence has to be established by playing. Human review remains the expensive step, which is why the maintainability complaints in the DK64 dispute matter: a port nobody understands is hard to fix when it breaks.
What is not known yet
The Verge's article is the only account of Sharma's note we could read, and we found no separate Xbox newsroom post. There are no announced products, dates, team size, or budget for Xbox Labs, and no stated policy on third-party ports of Microsoft-owned games. We will update this post if Microsoft publishes more detail.
Related reading
-
How AI agents actually decompile games — what the method can and cannot do
Details reflect reporting from October 13, 2026 and may change.