The phrase DeepMind EVE long-horizon agents mixes a real partnership with an unproven capability. On 6 May 2026 Fenris Creations, formerly CCP Games, announced independence from Pearl Abyss together with a Google DeepMind research partnership and a Google minority stake. On 21 August 2026 DeepMind published a games-research essay that casts EVE Online, a single-shard space MMO running since 2003, as the next hard environment after Atari, Go, StarCraft, and SIMA, because it never resets and because markets, alliances, and grudges unfold over years.
That environment choice matters for agent builders. DeepMind is publicly targeting continual learning, memory beyond context windows, planning over weeks to years, and multi-agent economics and diplomacy, the same failure modes that show up in production agent stacks. The public record through 23 October 2026 supports the staging and the problem statement. It does not support a claim that DeepMind has an agent that already plans on those timescales inside New Eden.
What we checked
- Has DeepMind demonstrated agents that plan over weeks or months inside EVE Online? refuted. Fenris and DeepMind describe an early-stage partnership that begins on a local, offline EVE instance using simulation and historical information. Neither the August 2026 DeepMind games essay nor Fenris's player letter publishes an EVE agent architecture, scores, or a multi-week planning evaluation. DeepMind's SIMA 2 paper still lists very long-horizon multi-step tasks and short interaction memory as open limitations. [6][8][11]
- Are autonomous DeepMind agents already playing in live EVE Online? refuted. Fenris's May CEO letter and DeepMind's August blog both state that initial research is not connected to Tranquility and that live EVE Online or EVE Vanguard would be considered only after capabilities mature. The only named player-facing system is Aura Guidance, a retrieval-based new-player helper, not an embodied game-playing agent. Fenris says it does not want systems that play EVE on a player's behalf or remove consequence. [9][6][8][13]
What we found
- The corporate deal and the research deal are real and distinct. Fenris's 6 May 2026 release says Google took a minority stake and that DeepMind will test models on an offline EVE instance running on a local server. Ashurst confirmed it advised Google on the equity investment and DeepMind on the partnership agreement. Pearl Abyss's 30 April 2026 KIND filing values the studio sale at KRW 177.132 billion, described as USD 100 million cash plus USD 20 million in token-acquisition rights, which Fenris cites as a USD 120 million transaction; that figure is not the size of Google's stake. The Verge reported that Fenris's CEO told Bloomberg the minority investment is in the millions; Fenris's own release does not state a dollar amount. [1][2][3][4][5]
- DeepMind's 21 August 2026 blog is a research agenda, not a results paper. It lists continual learning, memory beyond today's context windows, long-horizon planning over weeks, months, or years, and complex multi-agent dynamics as the reasons EVE is interesting, and it sequences the work as an offline EVE sandbox, then EVE Frontier for human-agent coexistence, then possible live EVE Online and EVE Vanguard only when capabilities are mature. It also flags EVE Vanguard's first-person tactics and EVE Frontier's programmable Smart Assemblies as distinct research surfaces. No EVE architecture, dataset, or success metric is given. Official Frontier builder docs do confirm that Smart Gates, Storage Units, and Turrets can be extended with player-deployed Move contracts. [6][7]
- Fenris's 20 August 2026 player letter is more conservative than the DeepMind essay. It says the work uses simulation and historical information on a local server, is at an early stage, and makes no promises about where it will lead. Fenris does not want systems that play EVE for players, sit above the rules, or remove risk. Suggested later uses include testing complex systems, tracing how changes spread through a simulated economy, helping new pilots, and possibly semi-autonomous governance systems. Tranquility is treated as 23 years of player history that should not be the first experiment surface; an opt-in research test server is mentioned as an example, not a scheduled product. [8][9]
- SIMA 2 is the public embodied-agent stack DeepMind places next to this work, and it is not an EVE result. The 13 November 2025 blog and the December 2025 paper describe a Gemini Flash-Lite agent that sees 720p frames and emits keyboard and mouse actions, roughly doubling SIMA 1 success on an expanded eval suite and generalizing to held-out ASKA and MineDojo, with self-improvement loops in ASKA and Genie 3 worlds. Both the blog and paper say the agent still struggles with very long-horizon multi-step tasks and must keep a short interaction memory for low latency. Fenris is absent from the SIMA 2 partner acknowledgements; the listed training games include No Man's Sky, Valheim, and Hydroneer, not EVE. The original 2024 SIMA agent was trained on tasks completable in about 10 seconds. [10][11][12][6]
- Aura Guidance is the only named live feature, and it is a retrieval assistant rather than a game-playing agent. Fenris's 17 February 2026 post describes a prototype that embeds a player's question, matches a bank built from Rookie Help (5.8 million messages analyzed from January through September 2025; tens of thousands of stored Q&A pairs), synthesizes an answer with limited character context, and refuses fleet, market, and fitting advice. The player's raw question is not sent to a language model; Fenris says there is no open-ended generation and that it runs small language-model variants, with an A/B test on new-player retention. DeepMind's August blog nonetheless says Aura uses Gemini. That attribution is DeepMind's, not Fenris's February write-up. [13][6]
What it means
Builders should treat this as DeepMind publicly naming the unsolved stack, continual learning, long memory, week-scale planning, and multi-agent economics, and buying access to a unique environment, not as a benchmark you can cite or a recipe you can copy. If you are designing agents, the transferable lesson is the staging: isolate a replay or offline world, measure forgetting and plan durability there, then consider opt-in human coexistence, and only then a live product. Do not assume a SIMA-style pixel-and-keyboard generalist will handle EVE's incomplete information, betrayal, and delayed payoffs; SIMA 2's own authors say long-horizon tasks and short memory remain unsolved.
For product AI, Fenris's Aura design is the more concrete pattern: retrieval over a vetted corpus, confidence gates, no tactical advice, A/B on retention rather than spend, and a stated willingness to shut the feature down. Copy that caution, not DeepMind's Gemini branding, unless you have equivalent retrieval and refusal constraints in writing. If you later put agents into a persistent economy, Fenris's player letter is the right product test: does the system create more meaningful decisions, or does it remove consequence?
Still open
- What agent architecture, training-data rights, and historical-replay fidelity DeepMind is using in the offline EVE instance
- How week-to-year planning, continual learning, and multi-agent behavior will be scored, if at all
- Whether SIMA 2, a Gemini hierarchy, or a different stack will be the EVE agent
- Timeline, consent, and disclosure rules for EVE Frontier human-agent experiments