EAIR paper finds enterprise agent teams break rules together

A new academic framework called EAIR (Enterprise Agent Interaction Rules), published on arXiv, targets a gap in how enterprise teams design systems where many AI agents work together. Individual agents may each stay within their own limits, yet the group still violates a business rule. No single agent looks wrong. The problem only appears at the level of the whole team. The EAIR researchers built executable test suites to check this failure mode. A company might set a spending cap or a data-access boundary. Each agent respects its own slice. Together, they exceed the cap or cross the boundary. Existing testing tools miss this because they check one agent at a time. EAIR checks the team. For a product team, the design question this raises is: who sees what the whole team has done, and at what point? A human reviewer looking at one agent's output sees a clean result. The problem lives in the sum. That gap between what one agent shows a reviewer and what the whole team has done is where approvals, budget limits, and access controls fail without the human noticing.