DeepMind is testing agents in an offline EVE sandbox, not on live servers

What we checked

• Is the DeepMind–Fenris program a staged, offline-first research partnership rather than live-server agents playing against EVE players? confirmed. Fenris, EVE Online, and DeepMind all state that work begins on an offline EVE Online instance on a local server, using simulation and historical information and not connected to live Tranquility. DeepMind then names EVE Frontier as the next human–agent coexistence testbed and says live EVE Online and EVE Vanguard would be considered only after capabilities are mature. [3][9][4][8]
• Is EVE being used because current generalist game agents still lack continual learning, long-horizon memory, and persistent multi-agent dynamics? partly confirmed. DeepMind lists continual learning, memory beyond today’s context windows, planning over weeks to years, and complex multi-agent economics as the reason EVE is a research opportunity. SIMA 2’s November 2025 post independently admits remaining failures on very long-horizon tasks and short memory bounded by a limited context window. No public agent results from the EVE sandbox have been released, so the mapping from those gaps onto New Eden is a stated program, not a measured result. [3][12][9]

What we found

• DeepMind’s historical claims check out against primary papers and blogs. The 2015 Nature DQN paper describes one algorithm, architecture, and hyperparameter set learning 49 Atari 2600 games from pixels and score. DeepMind’s October 2019 AlphaStar post reports Grandmaster on Battle.net for Protoss, Terran, and Zerg, ranked above 99.8% of active players, using a camera interface and action-rate limits. The August 2026 post uses that lineage to argue the field has moved from mastering a scored game to understanding a world without a score. [1][2][3]
• The commercial wrapper is an equity-backed studio partnership, not a public benchmark. Fenris’s 6 May 2026 announcement pairs independence from Pearl Abyss with a DeepMind research partnership on long-horizon planning, memory, and continual learning, plus a Google minority stake. The KIND filing matches Fenris’s USD 120 million figure as cash plus non-cash token rights. The Verge, citing Fenris’s CEO speaking to Bloomberg, reports the related investment as “in the millions” of dollars. Google’s exact ownership percentage is not in the sources opened here. [4][5][6][7]
• Staging is consistent across three first-party posts. EVE’s 6 May letter said initial research would be in controlled, offline versions not connected to Tranquility. EVE’s 20 August 2026 follow-up, after a joint knowledge-share, repeated the local-server start and said the work uses simulation and historical information. DeepMind’s 21 August post then published the longer ladder: offline EVE Online, then EVE Frontier, and only later possible enrichment of live EVE Online and first-person EVE Vanguard. [8][9][3]
• The agent stack DeepMind is generalizing is SIMA, not AlphaStar. The March 2024 SIMA report and blog describe an instructable agent that sees screen pixels, takes language instructions, and emits keyboard and mouse actions, with no game API or source access; early evaluation focused on skills completable in about 10 seconds. SIMA 2 (13 November 2025) puts Gemini at the core, adds conversation and a self-improvement loop, and still lists long-horizon multi-step tasks, short memory, and precise low-level control as open problems. EVE’s non-resetting markets sit on those failure modes. [10][11][12]
• The only named player-facing artifact is Aura Guidance, and DeepMind’s Gemini credit is not corroborated by Fenris’s own launch post. CCP launched Aura Guidance on 17 February 2026, before the public DeepMind deal: a retrieval-first helper built from Rookie Help chat, with 5.8 million messages analyzed from January through September 2025 and 706,000 genuine questions. The player’s raw question is never sent to a language model; small model variants synthesize from a vetted Q&A bank; uncertain queries go to human Rookie Help. DeepMind’s August post says the system “uses Gemini.” Fenris’s Aura post never names Gemini or DeepMind. [13][3]
• Fenris has published a hard product constraint. The 20 August EVE post says the studio is not interested in systems that play EVE on a player’s behalf, sit above the rules, remove risk, or replace player decisions with automation. DeepMind separately flags EVE Frontier’s programmable Smart Assemblies as an environment whose rules can change. Official Frontier docs describe Smart Gates, Storage Units, and Turrets as player-extendable via custom Move contracts, a different research surface from a fixed StarCraft ladder. [9][3][14]

What it means

If you build agents, the conclusion is about evaluation design, not about integrating an EVE API. DeepMind is saying match-based mastery is no longer the frontier it wants to advertise, and that SIMA 2’s remaining gaps—horizon, memory, continual learning, other agents who can defect—are now being pursued inside a licensed persistent world. You will not get those tests from Atari, a 10-second instruction suite, or a resettable sim. You need non-resetting state, incomplete information, and timescales longer than a context window.

If you operate a game, simulator, or enterprise digital twin, the operating pattern is studio partnership plus an isolated replica, with equity as a possible lock-in, and a public promise that live users are last. Fenris’s line against decision-replacing automation is the right default for any high-stakes sandbox: retrieve and explain, do not play the game for the user. Do not ship “Gemini NPCs” on the back of this story. No live agent, no public metrics, and an unconfirmed Gemini attribution on the one shipping feature are not a product spec.

Still open

• No public metrics, methods, or agent architecture have been released from the offline EVE instance.
• Google’s exact minority stake, and whether the USD 20 million token-acquisition rights relate to EVE Frontier’s on-chain assemblies, are not specified in the sources opened here.
• Whether the EVE agents are SIMA 2, a Gemini-only planner, or a new stack is undisclosed.
• Fenris has not independently confirmed DeepMind’s claim that Aura Guidance uses Gemini.