Google DeepMind Brings SIMA 2 Agents Into EVE Online

Google DeepMind is expanding its game-based AI research by partnering with Fenris Creations, the independent studio behind the massive multiplayer game EVE Online. This collaboration represents the culmination of a 15-year research arc that began with a Deep Q-Network learning 49 Atari 2600 games in 2015. It progressed through AlphaGo's 2016 victory over Lee Sedol, AlphaZero, MuZero, and the 2019 StarCraft II Grandmaster agent AlphaStar, before laying the foundation for the Nobel Prize-winning AlphaFold. Now, the lab is shifting from mastering single games to developing generalist agents that can operate in complex, persistent virtual worlds.

The core technology for this initiative is SIMA 2, the Scalable Instructable Multiworld Agent. Introduced in November 2025 with a Gemini model at its core, SIMA 2 expands on its predecessor's 600 language-following skills by reasoning about goals, conversing with users, and learning through self-directed play. While SIMA 2 has closed much of the gap to human performance in training environments like ASKA and Minecraft, it still struggles with long-horizon tasks and short memory constraints.

The EVE universe, which has run continuously since 2003, offers a single-shard economy and player-driven politics to test these limitations. DeepMind plans to evaluate four key capabilities: continual learning, extended memory, long-horizon planning over weeks or months, and multi-agent dynamics. The research will span three environments: the persistent EVE Online, the tactical EVE Vanguard, and the programmable EVE Frontier. Hilmar Pétursson, CEO of Fenris Creations, stated the teams are "pushing into uncharted territory" regarding AI learning timescales.

The research follows a staged path, starting in an offline sandbox before moving to EVE Frontier and eventually live servers. A prototype system called Aura Guidance, launched on February 17, 2026, is already live in an A/B test, using Gemini to answer rookie player questions. For AI practitioners, this partnership signals a transition from static benchmark testing to evaluating agents in dynamic, long-term social and economic systems, forcing models to evolve beyond short-context limitations.