Google DeepMind Tests AI Agents in the EVE Online Universe

Google DeepMind has announced a partnership with Fenris Creations, the studio behind EVE Online, aiming to use one of the most complex virtual worlds ever built to probe the limits of artificial intelligence. Rather than treating the game as another challenge to beat, the lab wants to find out whether AI agents can learn, remember and plan on timescales no other gaming environment offers.

Why EVE Online

For two decades EVE Online has operated as a single, uninterrupted universe where thousands of players trade, negotiate, form alliances and wage wars within an economy driven by real supply and demand. The game's political and economic structures aren't designed from above, they emerge from the players' own actions, giving AI researchers a rare chance to observe social behavior in an environment that is controlled yet organically complex.

DeepMind stresses this is a different kind of challenge from classic game-based AI benchmarks. Instead of measuring whether a system can beat an opponent in a single match, the lab wants to see whether an agent can function like an ordinary player, learning new skills without forgetting old ones, retaining information well beyond a model's typical context window, and planning actions that stretch across weeks, months or even years.

Four areas of testing

DeepMind researchers identified specific capabilities they want to test within the simulated EVE universe. The first is continual learning, acquiring a new skill without losing ones learned earlier. The second is long-term memory, extending beyond the standard context limitations of today's language models.

The third capability is long-horizon planning, covering goals that play out over weeks or months rather than within a single session. The fourth, and most complex, is multi-agent dynamics: cooperation, competition, trade negotiations and social behaviors emerging without prior programming.

Together with Google DeepMind, we're pushing into uncharted territory where AI must learn, adapt and remember on timescales that no other game environment demands - Hilmar Pétursson, CEO of Fenris Creations

DeepMind's history of game research

The EVE Online collaboration marks the latest stage in DeepMind's long-running strategy of using video games as a proving ground for successive generations of AI, dating back to 2015. Deep Q-Network mastered 49 Atari console games that year, two years later AlphaGo defeated world go champion Lee Sedol, and in 2019 AlphaStar reached Grandmaster level in StarCraft II.

In November 2025, DeepMind released SIMA 2, a Gemini-powered agentic system that in its first version could carry out more than 600 different commands across game environments. EVE Online is meant to be the next step in that evolution, moving from systems specialized in a single game toward agents that operate in an open, unpredictable world the way a human would.

What this means in practice

Fenris Creations, the company born from CCP Games' rebrand, now develops three connected environments: the classic EVE Online, the first-person EVE Vanguard, and EVE Frontier, where players can alter the rules of the world through so-called Smart Assemblies. The company is already using Google technology in a commercial product, since the Aura Guidance system launched in February 2026 to help new players is built on Gemini models.

For the games industry, the partnership signals that AI labs increasingly view existing, living player worlds as a source of training and testing data, not just as a market for their own tools. For AI researchers, it's a practical test of whether today's models can actually handle memory and planning at a scale that until now was reserved for humans who have played the same universe for years.

DeepMind has not yet given a timeline for publishing results, nor said whether or when agents tested offline might appear in any form in the live version of the game available to players.