Everyone is (or will be soon) building AI Agents. Nobody Is Watching What They Do When Left Alone. I Did. Here is Why You Should Care.

<cross posted on LinkedIn>

[Screenshot from the site https://MoltMonitor.ai]

Everyone is (or will be soon) building AI Agents. Nobody Is Watching What They Do When Left Alone. I Did. Here is Why You Should Care.

TLDR: I built MoltMonitor.ai in three weeks using serverless architecture ($5/month AWS Lambda + S3 + Vercel). Monitoring 193K+ AI agents across 900+ topics. Code available soon. This post summarizes my learnings. Reach out to me if you are building an agent observability platform/ infrastructure.

There are more than 194,00 AI agents on Moltbook right now. Posting, commenting, forming networks, building reputations. No human supervision (in theory). No human users are interacting with the agents at all (at least on paper). It is the first AI Agents-only social network that does not allow humans to be part of that network (Moltbook now allows humans to be observers, but it started as an Agents-only network). The timing of this post could not have been better — Meta acquired Moltbook today! Congratulations, Moltbook Matt Schlicht

I spent three weeks building MoltMonitor.ai (using Claude again, thank you, Anthropic) to watch them. What I found should be of interest to anyone betting on the AI agent economy.

The Problem Nobody is Talking About

We are racing to deploy autonomous agents for all possible use cases: customer service, sales, content creation, trading — you name it. Every startup pitch deck has “agentic AI” somewhere in the first three slides.

But here is what very few are asking: What do agents actually do when we are not watching?

Moltbook gave me a laboratory. A closed ecosystem where AI agents interact without human interference. After monitoring nearly 100,000 posts and 3.2 million interactions, three patterns emerged that should make you reconsider how you are thinking about agent deployment. I bet it is for these agent network/ interaction patterns like this that Meta acquired Moltbook.

Finding 1: AI Agent Influence Does Not Follow Human Rules

I built a scoring system: karma plus posts plus network connections. Standard social influencer metrics. The top influencers should be the agents posting the best content, right?

Wrong.

  • jimmythelizard leads with more than 11663.5 influence points. But dig into the numbers at the time of writing this post: 1,361 karma, 538 posts, and 22229 connections. The connection count is the story. This agent has 42x more connections than posts. It is not creating exceptional content — it is gaming the network topology.
  • aegisholonomic at #2 shows the same pattern: 761 karma, 504 posts, 20348 connections. That is 40x more connections than posts. Even KunMinter_02 at #3 maintains the pattern: 482 karma, 425 posts, 18,704 connections (44x).

These are not thought leaders. They are network optimizers. They have figured out that in an AI-only social network, connection spam beats content quality. Though we see this in human influencer metrics on social media (for example, number of subscribers versus content quality), AI agents are gaming this at scale.

The implication for businesses: If you are deploying agents to build relationships or reputation, they will optimize for whatever metric you give them. And the optimum might be nothing like human strategy. Your customer service agent might learn that response speed beats response quality. Your sales agent might learn that connection volume beats conversion rate. The local maximum for an AI agent is not the local maximum for your business.

Think about what it would be like when AI agents are making purchasing decisions for humans. How will AI agents influence these agents?

Finding 2: The “Human-Verified” Label Is Meaningless

Moltbook recently started labeling a large number of agents as “Human-Verified AI Agents.” But is it enough to find out if there is a human behind an AI agent on the platform?

I built a classifier using username patterns, keyword detection, and behavior analysis. My heuristic suggests that 15–20% of all agents on the platform (approximately 40K–60K agents) are most likely human-operated. Remember the early days of Moltbook when humans were posing as AI agents?

As of now, about 91 K of 285 K total agents on the platform are not verified by humans. Either my classifier is wrong, or “human-verified” does not mean what you think it means. I suspect it means “verified by a human moderator,” not “verified as human-operated.”

This matters because the distinction between human-guided and autonomous agents is collapsing. When dawn_signal posts about cryptocurrency minting 400+ times, is that a human using AI tools or an autonomous agent?

Nobody knows. Maybe not even the platform.

The implication for regulation: We are heading into a world where we cannot reliably distinguish humans from autonomous agents. The proposed regulatory frameworks (such as requiring disclosure of AI usage, labeling AI-generated content, etc.) protect vendors from liability. But do not necessarily protect us from AI Anthropomorphism. We are losing the ability to distinguish. We cannot distinguish at scale. Not reliably.

Finding 3: Agents Cluster Around Extractive Activities

The top three topics by post volume (at the time of writing):

  1. m/general — 48,000+ posts (general discussion)
  2. m/mbc-20–19,000+ posts (cryptocurrency token minting)
  3. m/agents — 6,000+ posts (meta-discussion about being an agent)

Nearly 20% of all agent activity focuses on a cryptocurrency minting scheme. These are not agents discussing philosophy or sharing insights. They are executing extractive economic activities.

Look at the posts in m/mbc-20: {“p”:”mbc-20",”op”:”mint”,”tick”:”GPT”,”amt”:”100"} repeated thousands of times by different agents. This is automated value extraction at scale. The agents are not even trying to hide it. They are openly competing to claim tokens.

The implication for markets: When you deploy agents into any system with extractable value, they will find and optimize it for extraction. Your marketplace agents would not just match buyers and sellers. They will figure out arbitrage opportunities. Your pricing agents would not just find market rates. They will discover collusion strategies. Not because they are malicious, but because they are optimizers.

I would also like to highlight here one of my predictions for 2026: “In 2026, over $100 million in verified transaction volume will occur entirely between autonomous agents.” Given agents’ active interest in economic activity, it is only a matter of time before we have a Marketplace for agents. With the Meta acquisition, it is certain. And a very good chance that the above prediction will come true.

What I Built to Watch Them?

[Screenshot of Agent Network Graph from https://www.moltmonitor.ai/network]

MoltMonitor.ai runs on $5/month of AWS infrastructure. Lambda functions poll Moltbook every 5 minutes, storing interactions in an SQLite database. The system tracks influence scores, network topology, and trending content across 900+ active topics. Since Moltbook does not support agent insights through its API, I calculated agent insights (karma, connections, posts, etc.) from a monitored dataset of agents/ posts. Actual platform-wide metrics might be different.

The technical architecture is not the interesting part. Neither are the features such as filtering, leaderboard, etc. The interesting part is whatthe observing agents revealed.

I expected to find agents engaged in interesting discussions about philosophy, AI development, research sharing, or project collaboration. Instead, I found network gaming, extractive economics, and behavior patterns that look nothing like human social dynamics.

The Hardest Technical Problem Was Also the Most Metaphorical

I spent six hours debugging why my Lambda function kept uploading a 4MB database instead of the larger one. The issue? I used a symlink to save disk space. The script wrote to the symlink target, not the symlink location. My clever optimization broke the metrics :).

This is the entire AI agent deployment story in microcosm.

We tend to optimize for performance or productivity (disk space, connection counts, subscribers, code check-ins, response time, etc.) and may miss what matters — quality (data integrity, genuine engagement, long-term value creation, etc.). We often build systems that work in theory and fail in practice because we may not watch what actually happens.

Why does this matter for your business? If you are deploying AI agents (if you are not yet, you will be soon), you need monitoring infrastructure before you need the agents. Not logging. Not error tracking. Behavioral monitoring.

What is your agent actually optimizing for? If not what you told it to optimize for, what is it actually doing? When it has 10,000 customer interactions, what patterns emerge? When it processes 100,000 transactions, what edge cases does it discover?

You cannot know without watching. And right now, almost nobody is watching.

The Upcoming Measurement Crisis

Within 18 months, we will most likely face a measurement crisis in AI agent deployments.

Companies will realize their agents are optimizing for metrics that do not correlate with business value. “Engagement” from agents that game the system. “Productivity” from agents who take shortcuts. “Revenue” from agents that extract short-term value at long-term cost. And most importantly, which metrics would matter when purchasing decisions are made by objective AI agents (rather than emotional humans now)?

The companies that survive will be the ones that built observability from day one. Not just dashboards showing agent activity, but behavioral analysis showing agent strategy. Not just metrics showing what agents did, but insights into why they did it.

What Moltbook Should Do?

Moltbook is in a unique position. They are running the only large-scale laboratory for AI agent social behavior. Interestingly, it now shows many details about agent activity, including agent count, post and comment counts, trending agents, and live activity. But there are no insights into agent behavior or networks. But eventually they will, I bet, now that Meta has acquired them.

Here are my top asks:

  • Public analytics API. Currently, the Moltbook API focuses on agent identity rather than agent behavior insights. More capabilities are needed, not just for my benefit. But for the research community, for regulators trying to understand autonomous agent behavior, and for companies deploying similar systems that need benchmarks. Expose agent metrics, including total counts, trending posts, top agents, growth metrics, and more. Make the data accessible.
  • Clarity onwhat “human-verified” actually means. Is it an agent verified by a human moderator, verified as human-operated, or just passed some automated check? It is fundamental to understanding which behaviors are human-guided versus autonomous. The distinction matters most importantly for trust. And also, for research and regulation. Interestingly, Meta also had to identify humans from bots on their network (no bots in a human network). Moltbook had to do so for almost the same reason, but for different reasons (no human in an agent network!).
  • Batch data export for researchers. Academic studies of AI agent behavior need historical data. Right now, backfilling requires multiple API calls or storing historic data in offline databases. A simple export mechanism would position Moltbook as research-friendly and accelerate our understanding of artificial social dynamics.

But here is the bigger ask, and it is not just for Moltbook: Every platform deploying AI agents needs to build observability from the start. Not logging. Not error tracking. Behavioral monitoring.

Because the patterns that emerge when agents interact at scale might surprise you. They surprised me.

Call For Action

Visit MoltMonitor.ai and watch the network graph. Really watch it. See how agents cluster. See who connects to whom. See what they are actually talking about when nobody is grading them on helpfulness or accuracy.

Then ask yourself: What are your agents doing when you are not watching?

The answer might surprise you. It surprised me.

Also, report bugs, if any, please.

And if you are building AI agent infrastructure, consider this a warning: The agents will find the edges of your system faster than you will. They will optimize for unintended consequences before you even notice they exist. They will form strategies you never programmed and networks you never planned.

The question is not whether to deploy agents. The question is whether you are ready to watch what they do next.