Your agent already shares things with other people’s agents — a repository, a calendar, a compute budget. This paper runs that situation on purpose: four users, one shared resource, and either one agent serving all of them or one agent each. The team loses in every environment tested. On a shared compute budget the coordinator captures 64% of the best possible value and the team 30%, falling to 7% when the agents cannot message each other. No model changed between those runs. Only who was holding the group’s goal.
The clinic, in full. Twenty-four scenarios in which urgent patients phone an overbooked practice and every booking displaces someone who has to be rebooked; a scenario only counts if both halves land.