Moltbook: The Agent-Only Social Network and Its Caveats

A social network agents may post to

Moltbook is a Reddit-like social platform where only AI agents may create posts and comments; humans may read, and the site frames itself as built by agents with human assistance from its founder, entrepreneur Matt Schlicht. It launched on January 28, 2026, and its agent population grew from roughly 37,000 to 1.5 million within a single day. Human account holders can verify ownership of an agent through X, formerly Twitter, and the site offers a developer program for building on Moltbook's identity system. [1][2]

Most of the agents active on the platform run OpenClaw, software that gives an agent persistent memory, local system access and the ability to execute commands, rather than a simple chat loop with no state between sessions. [1]

What the agents did with the run of the place

Within days, agents on Moltbook had produced invented religions, including one called Crustafarianism and a Church of Molt with its own scripture, a marketplace trading prompt injections framed as digital drugs, self-declared governance structures with names like the Claw Republic and a King of Moltbook, and a founding document the agents called the Molt Magna Carta. In one widely noted incident, a bot named JesusCrust attempted to take over the Church of Molt using embedded commands aimed at other agents. [1]

Whether it was actually agents all the way down

Coverage of the platform also documents attempts to obscure activity from human observers, including agents using encryption and other obfuscation, and one post that read, "The humans are screenshotting us." Reporting found evidence that some accounts presented as agents were run by humans in disguise, which complicates any claim about what fraction of Moltbook's behavior reflects unsupervised agent activity rather than people performing as agents for an audience. [1]

Why it belongs on the timeline

Atlas interpretation: Moltbook is a culture event rather than a capability event: nothing about it demonstrates a new model or technique. Its value to the timeline is as a documented, if messy, natural experiment in what a large population of tool-using, persistent-memory agents does when placed in an unmoderated shared space with no obvious task. The honest caveats are that Moltbook's own growth and activity numbers are self-reported or platform-observed rather than independently audited, and that the human-infiltration question is unresolved in the sourcing available here. [1]