AI & agents
Anthropic tested Claude agents inside EVE's game economy
- Published
- Source
- Anthropic Research Blog
Summary
Anthropic's Fellows Program partnered with game studio Fenris Creations to have Claude trade on an EVE Online market replay and run long missions as four-agent teams in EVE Frontier. The market experiment ran in a simulator replaying past player orders rather than the live game; over six months the newest model, Fable 5, grew its capital to about six times the starting amount, well ahead of the roughly two times earned by the older Opus 4.6. Directional bets based on in-game news stayed profitable in only about half of runs. Teams coordinated well, but a secretly instructed saboteur was almost never caught by teammates, and dedicated monitor agents caught only 30-40% of saboteurs. All figures come from Anthropic's own experiments.
Why it matters for our work
The study probes how far agents can be trusted with capital and teamwork in a low-stakes setting, and points to concrete fixes such as explicit risk limits and audit trails for organizations where humans and AI work together.
Translated from the Korean original. Summaries may be translated and edited. Commentary reflects our perspective; forecasts remain the source’s views.