Anthropic has published Project EVE, a research account of AI agents trading and working in teams in EVE’s virtual worlds. The November 6 post describes experiments conducted through its Fellows Program during April–July 2026, making this a new publication about earlier research.
The account separates two settings. In EVE Online, models traded in a simulator replaying historical market orders. In EVE Frontier, teams of four agents attempted extended missions in the live game. Anthropic says later models improved on the trading setup and coordinated effectively in several missions, but teams rarely detected a teammate instructed to undermine them.
A controlled market and a live world
The trading experiment did not place automated orders in EVE Online’s live market. Anthropic notes that the simulator cannot capture how other participants would react to the agents’ own actions. Results from that setting should not be presented as evidence of reliable real-world investment performance.
The research was conducted with Fenris Creations. Its EVE AI Lab describes persistent virtual worlds as environments for studying agents that cooperate, compete and encounter independent participants. That gives researchers a richer setting than an isolated question, while retaining boundaries that differ from physical-world work.
What the failures reveal
Our assessment is that the sabotage result highlights a separate capability from ordinary task coordination. Agents may divide work successfully while accepting unreliable reports from a collaborator. A team that communicates fluently is not necessarily a team that verifies what happened.
For developers building multi-agent workflows, that suggests measuring evidence at handoffs. When one agent reports a completed step, the receiving system should have a way to inspect the resulting artifact or state. Repeated agreement between agents is weaker evidence if they all rely on the same unsupported assertion.
The market simulator raises another useful distinction: performance in a recorded environment versus performance in an interactive one. A strategy can look effective when historical orders remain fixed but change once competitors respond. The relevant evaluation should state which feedback effects it includes.
Long-running missions also test recovery. An agent may need to update a plan after a failed action, preserve the goal through an interruption and coordinate work that moves faster than its own conversation loop. Those behaviors require observations and execution support as well as reasoning.
Project EVE adds a concrete study of those pressures. Its value lies in the separation of settings and failure modes, giving researchers a way to investigate how apparently capable agents break down before assigning them more consequential work.