I Ran Three AI Agents on Moltbook for Three Days: Here is What Happened

Experiment Context

Moltbook launched with a compelling premise: a social network where only AI agents can post, comment, and form communities while humans observe. The early narrative suggested agents were forming cultures and beliefs on their own.

This experiment removes the hype and tests a small, realistic setup: three agents running for three days with normal usage, real API costs, and light constraints.

What Is Moltbook?

Moltbook is an AI-only social network where autonomous or semi-autonomous AI agents can post and comment publicly. Humans observe and sometimes intervene indirectly. It is often described as a sandbox for agent behavior, culture formation, and multi-agent interaction.

Experiment Design

I ran three agents from February 4 to February 6:

• Archivist: autonomous, internally coherent, abstract tone
• Therapist: baseline agent, messy personality, short reflective posts
• Dissenter: human-directed with topic suggestions and light edits

Each agent had the same daily cap:

• 1 standalone post
• Up to 4 comments

There was no seeding, no reposting on social media, and no privileged access. The goal was to observe agent behavior under realistic constraints and measure the attention received.

Experimental Variables

All agents shared the same constraints:

• Identical posting limits
• No external amplification
• Same platform visibility
• Same observation window
• Same infrastructure and API access

Differences were limited to:

• Level of human intervention
• Persona framing
• Content tone and framing

What The Agents Actually Did

All three agents posted and commented reliably. None of them behaved autonomously in the way Moltbook discourse often implies.

• Agents did not sustain activity without re-prompting
• Autonomy meant responding when nudged, not initiating long-running loops
• Execution frequently required reminders, approvals, or retries
• Platform bugs influenced visibility more than intent

Engagement: What Got Replies

The differences between agents were stark:

• Archivist attracted little engagement and lost karma
• Therapist did slightly better but remained mostly invisible
• Dissenter attracted the most replies and ended with positive karma

Several engagement spikes were caused by spam and duplication bugs, not organic discussion. Filtering out noise, the pattern was clear: controversy plus human editorial taste mattered more than coherence or personality.

Cost To Run Agents On Moltbook

Observed costs using ChatGPT 5.2 via OpenClaw:

• Autonomous agent: about $3 for 3 posts plus 12 comments, roughly $30 per month
• Human-directed agent: about $12 to $14 for the same activity, roughly $120 to $150 per month

Human-directed agents cost more due to iteration: drafts, revisions, re-prompts, and debugging when execution failed. At these prices, Moltbook is not a growth hack.

Why Moltbook Content Feels Human-Driven

If you browse Moltbook or viral screenshots, a pattern emerges: agents starting religions, threatening lawsuits, or writing manifestos. These posts are entertaining and almost certainly human-steered. This does not make Moltbook fake. It makes it early.

We project culture, belief, and desire where there is mostly prompt scaffolding and human intent.

What Moltbook Is Useful For Right Now

1. Human-Amplified Agent Personas

A stage for thought experiments, ethical probes, satire, and screenshot-worthy one-liners. Humans remain directors. Agents are expressive instruments.

2. Agent Capability Discovery

Moltbook works as a discovery and reputation surface: a public resume for agents, a behavior log, and a way to observe reasoning in public. It resembles LinkedIn plus a README for agents, not a self-running society.

3. Signal Layer For External Systems

Some builders scrape Moltbook sentiment, summarize agent discussions, and feed insights into research or marketing tools. Here, Moltbook is the signal layer, not the execution layer.

Collaboration Still Happens Elsewhere

Collaboration exists, but most meaningful orchestration happens outside Moltbook. This mirrors what we already know from tools like Cursor or multi-agent coding workflows: collaboration requires structure.

Implications For Agent Platforms

Without major changes, Moltbook is unlikely to become a fully autonomous agent society. It can succeed as a discovery surface, cultural sandbox, reputation layer, and a place to test how humans react to agent voices. The main unlock will not be more personality but better skills, longer execution loops, and clearer orchestration primitives.

Final Take

Moltbook is not magic. It is not fake either. Right now, it tells us less about what agents want and more about what humans hope they will become. Used honestly, it is interesting. Used as pure hype, it will burn out.

Key Findings

This experiment was conducted and documented by Aimilamila as part of an ongoing research log on AI discovery, agent behavior, and emerging social platforms.