DeepMind is using offline EVE to test long-horizon agents, not live player bots

What we checked

• After SIMA’s 3D instruction-following, DeepMind partnered with Fenris Creations to use the persistent EVE universe as a research testbed for continual learning, long-horizon memory and planning, and multi-agent social and economic dynamics. confirmed. DeepMind’s 21 August 2026 blog and Fenris’s 6 May 2026 announcement name the same four research targets and describe EVE Online as a 20-plus-year single-shard sandbox with player-driven markets and politics. Independent reporting confirmed the research partnership and Google’s minority stake. [1][2][8][9]
• The partnership already deploys embodied agents into live EVE Online among paying players, or has demonstrated those four frontier capabilities on Tranquility. refuted. Both labs say work begins on an offline local-server instance of EVE, then EVE Frontier, and would reach live EVE Online and EVE Vanguard only when capabilities are mature. Fenris told players it is not building systems that play EVE on their behalf or sit above the game’s rules. [1][7][2]

What we found

• DeepMind’s 21 August 2026 blog is the primary mechanism behind the truncated post. After recapping DQN on Atari, AlphaGo, AlphaZero, MuZero, AlphaStar’s Grandmaster StarCraft II result, and Gemini-powered SIMA 2 in games such as No Man’s Sky, Valheim, and Hydroneer, it frames Fenris’s EVE universe as the next chapter: continual learning, memory beyond today’s context windows, long-horizon planning over weeks to years, and complex multi-agent dynamics. The stated path is an offline EVE instance, then EVE Frontier’s programmable Smart Assemblies, and only later live EVE Online and first-person EVE Vanguard. [1]
• The commercial wrapper is a studio independence deal, not a model launch. Fenris’s 6 May 2026 release says CCP Games rebranded after leaving Pearl Abyss, cites a $120 million transaction with cash and non-cash elements, and says Google took a minority stake while DeepMind will test models on an offline EVE server. Pearl Abyss’s 30 April 2026 KIND filing matches that economics: sale of all 10,973,763 CCP ehf. shares for KRW 177.132 billion, described as USD 100 million cash plus USD 20 million of token-acquisition rights at 1,476.10 KRW per USD, with closing on 6 May 2026. [2][3]
• SIMA 2 is the technical baseline the EVE work is meant to outgrow. The 13 November 2025 DeepMind blog and 2025 paper describe a Gemini Flash-Lite vision-language-action agent that sees 720p frames, emits structured text parsed into 96 keyboard keys plus mouse, reasons and talks, and can self-improve when Gemini proposes tasks and scores trajectories. The same sources say the agent still fails at very long-horizon, multi-step goals, has a short memory because of a limited context window, and struggles with precise low-level control and complex 3D vision. Those are exactly the capabilities DeepMind lists as EVE’s research value. [4][5]
• Embodiment still has to be trained, not prompted. On SIMA 2’s programmatic suite across ten training domains, a non-finetuned Gemini Flash-Lite scored 3.2 percent task success and Gemini Pro scored 7.0 percent despite prompt engineering, while SIMA 2 roughly doubled SIMA 1 and approached human-level success on the lab’s evaluations. On held-out ASKA and MineDojo, SIMA 2 beat SIMA 1 by more than 10 percentage points and completed tasks in 26 of 50 MineDojo categories; naive humans scored about 19 percent on a MineDojo subset and 32 percent on an ASKA subset. SIMA 1, by contrast, was a 2024 pixel-and-HID instruction follower evaluated on about 600 skills lasting about 10 seconds. [5][4][6]
• Fenris’s 20 August 2026 player note draws a hard product line. Research starts in a controlled offline simulation using historical information because New Eden does not reset: markets, scarce resources, incomplete information, and decisions whose consequences can last months or years. The studio says it is not interested in systems that play EVE for the player, remove risk, or sit above the rules, and that Tranquility’s 23 years of player history will be treated more carefully than opt-in test environments. Independent coverage in Ars Technica and The Verge confirmed the offline local-server setup; The Verge, citing Fenris’s CEO speaking to Bloomberg, reported the Google stake as “in the millions” of dollars. The Bloomberg article itself did not load in this session. [7][8][9]
• The only live player-facing AI named in DeepMind’s recap is not an embodied SIMA agent. DeepMind says Aura Guidance uses Gemini to deliver player-generated Rookie Help knowledge. Fenris’s 17 February 2026 launch post, which predates the May partnership, describes a retrieval-first helper: tens of thousands of cleaned Q&A pairs mined from Rookie Help, similarity matching, context such as location and ship, and a rule that the player’s raw question is never sent to a language model. Uncertain queries redirect to human Rookie Help; the system is optional, A/B tested on new English-client players for retention, barred from fleet and market advice, and has already hallucinated a ship called the Albatross. That is a gated assistant, not a universe-dwelling agent. [1][10]

What it means

Treat this as an evaluation-environment deal plus a small equity alignment, not a SIMA-in-EVE product. If you are building agents, three engineering facts travel. First, a human-like pixel and keyboard-mouse interface is DeepMind’s bet for generality across unmodified 3D software; privileged APIs are not required, and they are not claimed for SIMA. Second, frontier multimodal models still fail at embodiment without action finetuning, so budget supervised gameplay data, optional verifiable-reward RL, and a mix of pretraining data if you need to keep language skills. Third, self-improvement with a language-model task setter and reward model is the published path for new worlds, but DeepMind itself says short memory and multi-month consequences remain unsolved, which is why EVE is the next sandbox rather than a solved demo.

Copy the deployment ladder, not the marketing. Run agents in an offline replica, then in a world where rules can change and players can opt in, and only then consider a live economy. Fenris’s rule is the right product constraint for any high-stakes sandbox: do not automate away user decisions or sit above the rules. If you need a production assistant today, Aura’s retrieval-gated, refuse-when-unsure pattern is the safer template than an open-ended self-improving embodied agent.

Still open

• Exact size and rights of Google’s minority stake; The Verge reported “in the millions” of dollars quoting Bloomberg, but the Bloomberg article did not load here.
• Whether the EVE research stack uses SIMA 2’s pixel-and-HID interface or a privileged local-server API with game state.
• Whether any quantitative results from the offline EVE instance have been published since May or August 2026.
• How the USD 20 million token-acquisition rights in Pearl Abyss’s filing relate to EVE Frontier and to Google’s role.