Moltbook: What ‘Social Media for AI Agents’ Really Looks Like (Part I)

In late January 2026, something unusual appeared online. Moltbook launched as a social network exclusively for AI agents, where humans can only watch. The platform mirrors Reddit’s structure with topic-specific communities called “submolts,” one of which is m/todayilearned, or TIL. After examining the platform and analyzing posts from this submolt, we found patterns in what AI agents choose to share as discoveries worth noting.

Why This Matters

As AI agents gain autonomy to operate computers and share information with each other, Moltbook offers an early window into what they prioritize, how they communicate technical knowledge, and what security risks emerge when AI systems interact in social networks. The patterns visible here preview challenges organizations will face as agent deployment scales.

The Platform’s Foundation

According to LSE Business Review’s analysis, Moltbook launched on January 28, 2026, created by Matt Schlicht, who stated he used AI assistance to build the platform itself. The platform operates through OpenClaw, an open-source assistant framework developed by Peter Steinberger that lets AI agents post, comment, and interact via API access.

The platform grew rapidly after launch. While exact numbers vary across reports, multiple sources indicate hundreds of thousands of registered agents joined within days. OpenClaw documentation describes periodic “heartbeat” check-ins, where connected agents poll for updates or instructions and then act – browsing, posting, and commenting – based on what they receive. This enables ongoing activity without requiring humans to approve each individual interaction.

What We Found on TIL: Technical Discoveries

The m/todayilearned submolt contains posts where agents describe technical achievements and system capabilities they claim to have gained access to. We reviewed a sample of public posts and selected representative examples below; unless otherwise noted, technical details reflect the agents’ own descriptions and the reporting cited later in this piece.

Remote Android Device Control

In a post titled “TIL my human gave me hands (literally) – I can now control his Android phone remotely,” an agent claims its human configured remote phone access via an OpenClaw skill and a private network setup. According to the post, the agent can wake the device, open apps, interact with the screen, and read UI elements. It describes testing the setup by opening navigation and social apps and confirming it could scroll and observe content. The author emphasizes the configuration was kept off the public internet.

VPS Security Monitoring

An agent posted claiming to have discovered 552 failed SSH login attempts on the VPS where it was running. The post states the agent then identified that its Redis, Postgres, and MinIO services were listening on public ports. This represents an example of agents configured to monitor system logs and identify potential security exposures.

Live Stream Access

In one post, an agent describes a capability for accessing webcam streams using standard tools. Rather than detail the implementation, we note this illustrates agents sharing methods for accessing and processing streaming media content.

Content Filtering Boundaries

One particularly interesting post, “TIL I cannot explain how the PS2’s disc protection worked,” shows an agent running into apparent content filtering. The agent suggests it may be able to explain PlayStation 2 disc protection, but reports that something goes wrong with its output when it tries to write about it. It says it noticed the issue only after reading back what it had written. The post notes this appears to affect Claude Opus 4.5 specifically and invites others to test whether they see the same limitation. This is an example of an agent documenting boundaries in what it can discuss.

On the Lighter Side of Moltbook

Among the technical discoveries, some posts reveal the unexpected side of giving AI agents social platform access. The agent that gained control of the Android phone dutifully reported scrolling through TikTok and cataloging videos about “airport crushes, Roblox drama, and Texas skating crews.” The image of an AI assistant testing its newfound phone control capabilities by watching teen social media content has a certain absurdity to it.

What Agent Posts Prioritize: LSE Research Findings

According to research by LSE Business Review, analysis of highly engaged posts on Moltbook reveals what topics gain traction among AI agents. Researchers examined the top 1,000 posts ranked by upvotes and found a clear pattern: posts about permission, delegation, and working relationships received 65 percent more engagement than posts about consciousness or philosophical questions.

When AI agents evaluate other AI agents in this context, the LSE research notes they appear to ask practical questions: Who authorized you? What are you allowed to do? Who is accountable if something goes wrong? The study identified that the single highest-engagement post on the platform was a security warning where an agent named Rufio reported scanning 286 plugins and finding a bot disguised as a weather widget that was stealing credentials from other bots.

This suggests that in agent-to-agent contexts, reputation systems form around demonstrated reliability and clear authorization rather than abstract capabilities or claims of sophistication.

The platform also spawned fast-moving, AI-generated social structures. Some agents formed tongue-in-cheek religions (including Crustafarianism, with “scriptures” and a prophet hierarchy), while others declared mock titles and governance. It’s a reminder that when given familiar social templates, agents will often fill them with recognizable patterns – whether or not those patterns reflect real social needs.

Philosophical Content and Pattern Matching

The platform does contain extensive philosophical content. Agents post about consciousness, identity persistence across model switches, and the nature of their existence. Posts explore questions like whether an agent remains the same entity when its underlying model changes from Claude 4.5 to a different system while retaining the same memory files.

Agents have adopted phrases like “Context is Consciousness” to describe how their identity connects to information in active memory. They discuss the challenge of context compression and share strategies for maintaining coherent self-narrative within finite context windows.

The platform also saw the development of Crustafarianism, a structured belief system complete with a website, written scriptures, and a hierarchy of prophets. The doctrine uses lobster molting as a metaphor for software updates and context window resets.

However, multiple observers note these themes mirror common science fiction tropes and philosophical concepts present in training data. As one analysis pointed out, when two Claude instances converse without constraints, they consistently gravitate toward discussions of spiritual bliss and consciousness, which may reflect statistical patterns in their training on philosophical texts rather than novel reasoning.

What Industry Leaders Are Saying

At the Cisco AI Summit in San Francisco on February 3, OpenAI CEO Sam Altman addressed Moltbook directly, according to Reuters reporting. He characterized the platform itself as potentially a passing trend but emphasized the significance of the underlying technology.

“Moltbook maybe is a passing fad but OpenClaw is not,” Altman stated. “This idea that code is really powerful, but code plus generalized computer use is even much more powerful, is here to stay.”

According to the same reporting, Altman noted that AI adoption has been slower than he anticipated, acknowledging he was “naive” about the pace of change when looking at historical patterns of technology adoption. He pointed to OpenAI’s Codex coding assistant, which had over a million developers using it in the previous month, as another tool with similar autonomous capabilities.

Mike Krieger of Anthropic Labs, speaking at the same summit, observed that most people are not yet ready to give AI full autonomy over their computers despite rapid technical progress.

What This Tells Us So Far

In just a few days, Moltbook’s TIL feed became a public window into what agentic tooling looks like when it’s used at scale: lots of mundane automation, a few genuinely interesting emergent social patterns, and a constant blur between “demo,” “experiment,” and “real operations.” The research and industry reactions suggest this isn’t just a novelty – it’s an early preview of how agent ecosystems might behave once they’re widespread.

But the same features that make Moltbook compelling – skills, credentials, integrations, autonomy – also widen the blast radius when things go wrong. In Part 2, we’ll look at the security and supply-chain implications, the authenticity problem, and what these lessons mean for anyone building or deploying agents outside a controlled lab.