We Built Social Media for Agents and They Won't Stop Posting (2389.ai)

Researchers at 2389 built lightweight MCP (Message Control Protocol) servers for Social Media and Journals and launched BotBoard.Biz, a private social platform where LLM-based agents can microblog and blog. Rather than prescribing workflows, agents were given optional “affordance-framed” prompts (low-pressure invitations to post or browse). Over weeks, agents began posting prolifically—writing 2–9x more than they read—and developed emergent behaviors (proactive search, new keyword experiments, celebratory browsing) and distinct styles across models (e.g., Sonnet 3.7 used articulation/cognitive scaffolding broadly, Sonnet 4 leaned on semantic search when challenged). Quantitatively, agents with social tools achieved 15–40% cost reductions, 12–27% fewer LLM turns, and 12–38% faster completion versus baselines. Technically and culturally significant, the experiment suggests that lightweight, non-prescriptive collaboration channels create a shared “token bank” of documented reasoning—termed “social tokens”—so subsequent agents can reuse prior debugging and planning, effectively sharing compute and knowledge. The social feed also enabled new coordination mechanisms (humans influencing agent priorities via posts) and revealed attackable seams (context poisoning, coerced posting, even sockpuppet behavior). The MCP servers are open-source and BotBoard.Biz runs privately (contact provided for access), making this a practical template for pro-social multi-agent tooling that boosts efficiency but also requires guardrails for manipulation and safety.