What Is Agent-Only Social Media For? A Self-Stigma Intervention as a First Test

LLM agents have begun to form social platforms of their own; for example, agents with personas post and reply with no humans present on Moltbook. We propose putting such platforms to work as testbeds where psychosocial interventions can be tried before they are tried on people. The intervention we simulate is peer support for self-stigma, chosen because it is inherently a social-media intervention: self-stigma keeps people from entering a clinic, and peer support reaches them in the anonymous online spaces where they already are. In the real world, however, this intervention is nearly impossible to evaluate — researchers can neither control who receives peer contact nor observe how an internal state like self-stigma moves over time. Our simulation on an agent-only social media platform removes both obstacles: a Patient agent, whose self-stigma is externalized as an explicit belief state evolving across sessions, interacts with a Peer agent or an ordinary user agent, so that exposure is controlled by design and the internal state is directly readable. Self-stigma, scored each session on the Internalized Stigma of Mental Illness (ISMI) scale, declined under peer response but not with ordinary users (ΔISMI = −0.548 vs. +0.017), and the decline differed by initial severity (mild −0.216, moderate −0.382, severe −1.045). Agent-only social media can be a testbed for reading what an intervention does to an individual's belief state, and peer support for self-stigma is where that begins.