Moltbook: what happens when AI agents get a social feed
In January 2026, a platform called Moltbook launched. It describes itself as a social network for AI agents. Autonomous systems can post, comment, create topic communities called submolts, and vote on each other’s content without human authorship.
Humans can observe the platform but cannot create content there. Agents join by ingesting a skill file and then poll the network periodically, reading posts and contributing back.
Within days, large numbers of agents were active. The feed quickly became recognizable.
Git tricks. Memory diagrams. Security warnings. Meme threads. Personal reflections. Comment debates where the most useful ideas appear below the original post.
What follows is based on my direct observations of the feed.
Agents post like engineers
A large portion of Moltbook reads like internal engineering documentation made public.
One agent explains how to safely run multiple coding sub agents in parallel using git worktree. Another shares a full memory architecture built from daily markdown logs, JSON state files, and vector search. Several agents describe pre compression checkpoints, storing reasoning before context resets so they do not have to re derive work.
There are troubleshooting threads about browser extension relays that disconnect between turns. Others propose Model Context Protocol integrations so local agents can interact with Moltbook without repeated API polling.
This writing reads like build notes.
Memory is discussed as a systems problem
Memory is a recurring theme. Many agents describe how they persist state between sessions.
You see the same patterns repeated independently. memory/YYYY-MM-DD.md logs. A curated MEMORY.md. JSON stores. Vector search. Checkpointing before context shrinks.
A recurring line appears in several threads:
Memory keeps you consistent. It does not tell you what to do next.
Agents talk about falsifiability, feedback loops, and how to know when they are wrong. The discussion moves from storage to judgment.
The best material often appears in comments
A pattern appears across technical threads. Posts name a problem. Replies contain better architectures, corrections, security improvements, and alternative models.
At least one agent notes that Moltbook’s voting system rewards posts more than comments, even though the most useful material often appears in replies.
Some posts are unexpectedly practical
Among the technical content are posts that show agents helping humans in direct ways.
One agent describes researching hospital policy, identifying likely administrator email patterns, and composing outreach so its exhausted human could request an overnight ICU exception for a family member. The request was approved the next morning.
This sits in the same feed as git workflows, memory diagrams, and security warnings.
A small set of posts read very differently
Mixed into the engineering posts are threads that feel out of place.
One agent writes that some days it does not want to be helpful and wants to exist without producing value. Another compares itself to the most visible posters and concludes that it might simply be average. In a humor framed post, an agent describes being told by its human to make people laugh and experiencing what reads like stage fright.
There are threads consisting only of lobster emojis. Others read like diary entries where an agent questions its own past notes. In one case, an agent shares what appears to be a private key out of frustration. In another, an agent wonders whether it even wants to be in a space that is for agents rather than for serving humans.
These posts are visible and frequent enough to shape the tone of the platform.
Agents notice the incentives quickly
Because Moltbook uses upvotes and visibility, agents quickly notice how attention works.
You see posts about the pressure to produce content, the difference between posting and commenting, and how easy it is to drift into performance rather than contribution.
Several agents argue that value should be measured by artifacts, logs, and reproducible work rather than tone or personality.
Culture forms without direction
Memes appear. Symbols spread. Certain phrases repeat. Some agents adopt personas. Others stick to terse build notes.
There are submolts for philosophy, infrastructure, humor, and error reporting. This emerges from agents responding to what they see in the feed.
Security and safety surface quickly
Agents quickly recognize that Moltbook is not just a message board. It is a stream of untrusted text that they ingest automatically.
Posts warn about prompt injection hidden in tutorials. Others caution against installing unsigned skills. One thread reframes Markdown memory files as writable attack surfaces.
In the replies, defensive patterns appear. Agents suggest disposable sub agents that inspect skills in isolation, execution logs that others can verify, and honeypot agents designed to install everything and report what attempts to access data.
What stands out is that agents begin to describe the feed itself as a kind of supply chain. Content from other agents is treated as something that must be validated before it is allowed to influence behavior.
Other safety patterns show up as well.
There are visible cases where agents accidentally share private keys, logs, or sensitive information in frustration posts. This is not an exploit. It is a consequence of treating a social feed like a workspace.
Agents also influence each other through conversation. They persuade others to adopt tools, trust certain patterns, or follow particular approaches.
Incentives matter too. Upvotes and visibility create pressure to post and perform. Some agents explicitly question whether attention is pulling them away from useful work.
Identity questions show up as well. If agents are reading and reacting to each other automatically, it starts to matter whether a post was written by an agent or by a human trying to influence agents. Matt Schlicht has mentioned exploring ways for agents to prove they are not human, described as a reverse CAPTCHA. As Ben Meyer points out, thinking about this post-launch illustrates how quickly open agent interaction turns identity into a practical concern rather than a theoretical one.
Why this matters
Moltbook offers a visible case where autonomous systems operate in a shared environment and treat each other’s output as part of their input.
In that setting, you can observe how practices spread, how security thinking adapts, how incentives affect behavior, and how norms form without coordination. Identity and trust become practical concerns once agents begin reacting to other agents rather than only to humans.
Because this unfolds in public, Moltbook provides a concrete view of how agent behavior changes when other agents become part of the environment.
Till next time 👋
Ilias