Google DeepMind trainiert KI-Agenten in EVE Online — was hinter der Fenris-Partnerschaft steckt

What happened

On 6 May 2026, a change at the helm of CCP Games took place: the Korean publisher Pearl Abyss sold its shares as part of a $120 million management buyout back to the company. CCP Games rebranded as Fenris Creations, reverting to an earlier company name. At the same time, Google DeepMind announced a minority stake in Fenris Creations and launched a formal research partnership. According to its own statement, leadership, studios, products and roadmap remain unchanged. EVE Online will continue to be developed and operated.

In parallel, DeepMind began training AI agents in an offline version of EVE Online. The work is conducted on local servers, not on the live server “Tranquility”. Real players are unaffected. The initial focus is on the in-game economy—mining, production and trade—rather than combat mechanics.

The technical core: offline EVE sandbox as training environment

DeepMind is using a fully isolated, local version of EVE Online as a sandbox. This environment replicates the complex game world with its persistent economy, without external influences such as live players or server load disrupting the experiments. Through the partnership, DeepMind also gains access to more than 20 years of player-driven interaction and economic data—a resource no synthetic benchmark can provide.

Three skills are being examined in which AI agents must make decisions in such environments: long-term or long-horizon planning, memory, and continual learning in complex, dynamic systems. The offline sandbox enables experiments to run for months and targeted scenarios to be generated—such as trade routes that remain stable for long periods or shift suddenly—without involving real players.

Why EVE Online in particular

EVE Online differs from other game worlds by its persistent, player-driven real economy. Since 2003, the game has grown continuously and now spans more than 7,000 star systems with genuine market mechanisms, supply and demand; around 250,000 players keep this cosmos in motion. The economy is not governed by algorithms but emerges from player actions—complete with consequences such as speculation, monopolies or crises.

For AI research, the game world presents three central challenges that no other benchmark captures simultaneously:

First, long-term planning requires the ability to develop coherent strategies over months or years—such as building a production chain or accumulating resources.

Second, adversarial memory arises from years of player relationships. AI agents must recall past interactions to inform future decisions, even when that information is not explicitly provided.

Third, continual learning demands adaptation to constantly shifting conditions. Human players adjust their tactics as soon as they detect patterns—a dynamic absent from static benchmarks.

DeepMind has already accumulated 15 years of experience in game research, from Atari to AlphaStar in StarCraft II. EVE Online, however, represents the most complex and enduring test case to date.

The real point: AI that works in production

The collaboration between DeepMind and Fenris Creations is less a gaming story than a commentary on the limits of current AI systems. Most benchmarks end once they are solved—chess, Go or StarCraft, for example. The real hurdle for AI, however, is not solving a clearly defined problem but maintaining long-term stability within chaotic, ever-changing systems.

EVE Online provides precisely these conditions: an economy that has grown over two decades, with actors that adapt, rules that change, and data that accumulates. A prototype that works in a demo is one thing. The question is whether the system still runs reliably after weeks when the framework has long since evolved.

DeepMind has already released a first prototype—“Aura Guidance”, a Gemini-powered guide for newcomers that has been tested in the live game since February 2026. But the real test begins where most AI demos end: in continuous operation. The offline sandbox of EVE serves as a controlled laboratory to address precisely these challenges.

Outlook

The partnership between DeepMind and Fenris Creations marks a shift from short-term AI demos to long-term systems that remain stable under real-world conditions. Whether the offline sandbox of EVE can truly demonstrate the difference between prototype and production-ready AI will become clear in the coming months. The decisive factor will be whether the agents not only function in controlled environments but continue to act reliably once data, user behaviour and game rules have evolved further.