On August 21, 2026, Google DeepMind announced a research partnership with Fenris Creations, the studio behind EVE Online, alongside ten other game developers including Hello Games (No Man’s Sky), Coffee Stain Studios (Valheim, Satisfactory) and Foulball Hangover (Hydroneer). The post frames the collaboration as the continuation of fifteen years of DeepMind games research, the lineage that runs from Atari through AlphaStar and MuZero.
The agent at the centre of it is SIMA 2, the Gemini-powered Scalable Instructable Multiworld Agent, which DeepMind describes as capable of real-time reasoning and conversation across complex 3D game worlds. The stated research targets are four capabilities that generalise well past gaming: continual learning, memory, long-horizon planning, and complex multi-agent dynamics. EVE Online is the reason the partnership is interesting rather than promotional - it is a single persistent universe that has been running for more than twenty years with a real player economy, alliances and betrayals, which makes it one of the few available environments where all four of those capabilities are exercised at once and where the ground truth is a live population rather than a scripted scenario.
DeepMind also points at something already shipped: the Aura Guidance system in EVE Online, which uses Gemini to answer new-player questions from player-generated knowledge drawn from real Rookie Help exchanges. The lab describes its approach as “show, don’t tell” - building playable prototypes with developers rather than publishing purely theoretical results.
For a business or technical leader, the signal here is about where the hard problems in agents are being pushed, not about games. Continual learning and durable memory are precisely the capabilities enterprise agent deployments keep running into, and they are notoriously hard to evaluate in office-work settings because there is no clean environment and no cheap reset. Persistent multiplayer worlds are a plausible substitute: adversarial, economically consequential, and instrumented. What the announcement does not include is any published capability result, so this is an agenda and a set of environments rather than evidence that SIMA 2 has solved anything.