DeepMind Puts SIMA 2 AI Agents to Work in EVE Online

Games have been AI research infrastructure since DeepMind taught a network to play Atari in 2013. On August 21, 2026, the lab published a look back at fifteen years of that work and named its next environment: EVE Online, the space MMO that has been running continuously since 2003. The research partner is Fenris Creations, the studio that became independent from Pearl Abyss in a $120 million management buyout in May 2026.

• The collaboration spans EVE Online, the tactical shooter EVE Vanguard, and the programmable EVE Frontier
• Research runs on offline, controlled copies of the game on local servers, fully separated from Tranquility, EVE's live cluster
• Agents receive a stream of pixels from the game display rather than privileged access to internal game state
• SIMA 2, DeepMind's Gemini-powered generalist agent, is the system under test; other partner studios include Coffee Stain, Hello Games, Foulball Hangover, and Keen Software House

Why EVE Online Is a Harder Problem Than Chess

Board games end. EVE does not. It is a single shared universe with a player-run economy, multi-year alliances, industrial supply chains, and consequences that compound across real-world years. An agent that performs well there has to hold goals over horizons measured in weeks or longer, remember what happened, and reason about other agents whose incentives shift.

DeepMind names four capabilities it wants to stress: continual learning in a world that keeps changing, long-term memory, long-horizon planning, and multi-agent dynamics. Those are the same four gaps that show up whenever a language model agent is asked to run a long task, which is what makes this more than a gaming curiosity. The question of how much context an agent should carry is being probed from several directions at once — including recent work on calibrating agent memory to model capability.

How Does SIMA 2 Actually Play the Game?

SIMA 2 is the second generation of DeepMind's Scalable Instructable Multiworld Agent, introduced in November 2025 with a Gemini model at its core. The first version learned more than 600 language-following skills across commercial games; the second reasons about goals, holds a conversation with a human, and improves through self-directed play.

The constraint that makes the results meaningful is the input channel. SIMA 2 reads pixels from the rendered display, the same information a human player has, rather than querying the game engine. That is deliberately harder than the alternative, and it is why the approach transfers across titles without studios modifying their game code.

Running on offline copies also answers the obvious concern from EVE's player base. Nothing in this research touches the live economy or competes with players in Tranquility. Fenris has separately shipped Aura Guidance, which uses Gemini and player-written knowledge to help new pilots — the practical, shipped-today half of the relationship.

What Comes After Games

The through-line from Atari to EVE is that games are cheap, safe, repeatable environments for problems that are expensive and risky everywhere else. The same research on planning and world modeling shows up in robotics work like Gemini Robotics 2's whole-body control and in edge world models such as NVIDIA Cosmos 3.

A persistent universe with twenty-three years of accumulated player behavior is an unusually rich test bed, and one no simulator could have been built to imitate. More AI research coverage is on our artificial intelligence page.

Sources: Google DeepMind — August 21, 2026; Fenris Creations — 2026; Unite.AI — August 2026.