AI agents sharing a virtual space need traffic rules, and a new audit shows most of those rules are bad at letting anyone through.
Researchers built a typed "snapshot settlement" contract that arbitrates when multiple AI agents propose valid but conflicting actions in the same environment, then audited it for three properties: whether outcomes depend on event order, whether agents actually make progress, and whether action logs replay consistently. They ran 28,800 exhaustive permutation trials and 2,160 scripted multistep episodes against five different settlement policies. In a six-agent doorway task, a "conservative rejection" policy let only 31.25% of agents complete their moves, while a policy based on random tickets let 90.28% through - a 59-point gap. Every policy respected the spatial rules it was supposed to enforce, but even priority-based arbitration fell short of the best possible outcome for small test cases. A separate journal audit replayed 156 checkpoints exactly and correctly rejected 1,332 deliberately corrupted logs.
As companies deploy swarms of LLM agents to act concurrently in shared digital environments, the referee logic deciding whose action wins becomes as important as the agents' own reasoning. This is the same problem distributed databases and operating systems have wrestled with for decades - transaction isolation, lock contention, scheduling fairness - just rebranded for agent frameworks. The paper's numbers suggest the obvious safe choice, rejecting anything ambiguous, can be the worst one for actually getting work done.
The authors are careful to note this measures execution correctness in a toy doorway scenario, not whether any of this holds up once real agents are fighting over real-world resources.