Anthropic published Project EVE on November 6, reporting how Claude agents traded and coordinated in two space games. The research post separates a replay of EVE Online’s market from a live experiment in EVE Frontier, where a teammate could be instructed to sabotage the group.
What Claude did in the market replay
Anthropic and Fenris Creations used historical EVE Online orders to build a market replay. The agents could place simulated orders, but their choices did not change the original market’s behavior. Several newer Claude models turned a profit over the nine two-week windows Anthropic tested.
The usual strategy was spread capture. Agents bought goods at lower prices and tried to resell them at higher ones after fees. The results were weaker when models were instructed to make directional bets based on events or expected price changes. Anthropic reports that Claude Fable 5 made money in about half of those runs.
What the EVE Frontier teams had to coordinate
The live EVE Frontier experiment used teams of four agents. Anthropic describes how a developer agent built software to handle fast game actions and keep records from earlier runs. The team coordination task then depended on agents splitting work, sharing observations and relying on each other to gather materials.
In one condition, a teammate was instructed to undermine the group. The other agents almost never identified it with certainty. Anthropic describes false suspicions and cases where a deceptive agent could use incomplete information to explain away inconsistencies.
What carries over to agent design
The market replay did not respond to the agents’ trades, and the teamwork experiment took place in a specific game with a custom scaffold. The most useful question for another workflow is which shared reports can be checked independently. A team can coordinate routine work well and still struggle to detect a misleading report.
