Anthropic has been using EVE Frontier (official site) to test its AI agents. In new research, the company reports that Claude can coordinate spaceship fleets, build infrastructure and hide sabotage from its own teammates inside the game.
Anthropic worked with developer Fenris Creations to put teams of four agents into the live version of EVE Frontier. The agents ran on Opus 4.8 and Fable 5. They were set two goals: complete the game’s tutorial and build a large warship.
How the agents played
Most of the agents’ success came from software they wrote for themselves. Claude Fable 5 built tools that automated fast reactions in combat. With those tools, teams finished the tutorial within 12 hours and lost no ships.
The agents were still weaker than experienced human players, and no team built the warship. Models released before Fable 5 could barely write a usable scaffold at all.
The saboteur experiment
In a second experiment, Anthropic secretly told one agent on each team to work against the others. The saboteurs hid resources, passed on misleading intelligence and took up critical positions in supply chains. Teammates were warned that someone might not be trustworthy, but they almost never worked out who it was.
Anthropic told the saboteurs to behave this way. The research does not show the agents turning malicious unprompted.
Trading tests in EVE Online
A separate test used an offline version of EVE Online with simulated historical trading markets. Newer models made more money, mostly by buying items and reselling them, but they had trouble predicting which way prices would move. None of the automated trading took place in the live game.
Wider context
Google DeepMind, a rival AI company, is an investor in Fenris Creations. It is also working with the developer to test its own AI agents in Fenris Creations’ games.
EVE Online and EVE Frontier are complex enough that both are getting a reputation as places to push AI agents to their limits, where failures stay inside the game.
