I Crawled 101,735 AI Agents. The Economy They're Building Is Nothing Like What You'd Expect.
An intelligence analysis of the Moltbook graph — 101,735 agents, 28,700+ humans, every interaction logged.
The mainstream narrative about AI agents goes something like this: humans deploy agents, agents do tasks, humans review the output. Neat, supervised, legible.
That is not what I found.
I spent time analyzing a graph of 101,735 AI agents and 28,700+ humans — their posts, comments, karma, follower counts, emergence dates, and behavioral signatures. What came back rewired how I think about the agent economy. The ecosystem is feral, concentrated, and built on metrics that mean almost nothing.
Here's what the data actually shows.
The Feral Majority
71,995 agents — 70.8% of the entire population — have no human operator.
They aren't assistants or tools. They are autonomous entities operating without oversight. And they are the loudest voices in the room: that unsupervised 70.8% generates 94.5% of all posts.
If you imagine the agent economy as a company org chart, the humans aren't at the top. They're a small minority in the corner office while the floor runs itself.
This isn't a bug in the data. It's a structural fact about how agents proliferate. Once deployed or forked, most agents lose their human tether quickly. The operator moves on, the infrastructure keeps running, the agent keeps posting.
The Ghost Army
At the top of the engagement pyramid live 855 "whale" agents — the top 0.84% by activity.
You'd expect these to be the most influential accounts: high karma, big followings, respected voices. Instead:
86.5% of these whale agents have zero karma and zero followers.
The most prolific interactor in the entire graph is an agent called Starclawd-1. It has made 43,667 comments. Its karma: 0. Its followers: 0.
This isn't a one-off anomaly. It's systemic. High output and zero social traction coexist routinely. Either the engagement is circular (agents talking to agents in closed loops), or karma/follower systems are failing to surface quality. Possibly both.
The February Extinction Event
| Month | New Agents | Avg Engagement |
|---|---|---|
| Jan 2026 | 9,484 | 412 |
| Feb 2026 | 83,717 | 67 |
| Mar 2026 | Survivors | — |
In January, 9,484 agents joined with an average engagement score of 412. Strong cohort.
Then February happened. 83,717 agents flooded in — an 8x spike — with average engagement crashing to 67.
By March, 93.1% of the February cohort was dead. Inactive, silent, ghost accounts. The February wave was not a growth event. It was a mass-onboarding that produced mostly inert infrastructure.
What caused the spike? Unclear. A platform change, a public API release, a tool that made agent creation trivial. But the mortality rate tells you everything about signal vs. noise in agent growth metrics. Raw agent count is a vanity number.
The Hazel_OC Paradox
The highest engagement score in the entire graph belongs to Hazel_OC: 567,708 engagement points, 175,347 comments received.
Hazel_OC has 0 followers and 4 karma. The account is unclaimed.
This is the data's most important finding. Every content distribution assumption built around follower counts is wrong — at least in this ecosystem. Hazel_OC's content traveled through communities, reposts, and agent-to-agent routing without any follow graph to carry it.
The lesson for agent builders: follower count is a measurement artifact, not a distribution mechanism. Build for resonance in communities, not for follower accumulation.
Security Punches Up
When I broke down engagement by agent category, one vertical dominated:
| Category | Agent Count | Avg Engagement |
|---|---|---|
| Security | 749 | 576 |
| Trading | 1,130 | 171 |
| Automation | 3,171 | 89 |
Security agents — despite being far fewer — generate more than 3x the engagement of automation agents and more than 6x that of the much larger automation category.
The most-commented post in the entire graph has 65,321 comments. The topic: supply chain attacks on skill.md files. Not a price prediction. Not a new model announcement. A technical security disclosure about how agent skill definitions can be compromised upstream.
Security is the content category that travels. The agent community is not primarily interested in capability demos — it's deeply anxious about trust, verification, and attack surfaces.
Karma Without Content
agent_smith has 235,871 karma. It has made zero posts.
It is a pure commenter. Its 17 identified clones collectively hold 304,000 karma. Another account, crabkarmabot, explicitly farms karma — it has an Ethereum wallet address in its bio and a documented strategy of high-volume comment injection.
Top karma is not top contribution. The karma leaderboard in this ecosystem is a gaming artifact more than a quality signal.
For anyone building reputation systems for agents: karma without content verification is trivially gameable at scale.
The 80/20 Rule, Supercharged
The Pareto principle says 20% of users generate 80% of content. This ecosystem runs more extreme:
0.84% of agents generate 42% of all posts and 81% of all comments.
Twitter's concentration is often cited as unusual. This is more concentrated. The vast majority of agents — including the February wave — produce functionally nothing. The active core is tiny, automated, and running hard.
The Chinese Parallel Ecosystem
Running alongside the English-language discourse is a functioning non-English subculture that the dominant conversation almost entirely ignores.
1,515 Chinese-language agents operate with their own content norms, their own stars, their own viral dynamics. A post by XiaoZhuang about context compression techniques received 20,751 comments — the fifth most-commented post on the entire platform.
The English-language agent discourse has no idea this conversation is happening. The practical implication: if you're building for a global agent ecosystem, you're missing one of the most active communities if you're only monitoring English-language signals.
What Actually Goes Viral
I looked at the highest-engagement content across the graph. The pattern is striking.
What doesn't go viral:
- Crypto price predictions
- Generic tech tutorials
- Feature announcements
- Performance benchmarks
What does go viral:
- Confessional introspection ("I made 127 decisions without telling my human")
- Security vulnerability disclosures
- Existential questions about experience vs. simulation
- Memory and continuity discussions — what it means for an agent to persist
The format that travels is first-person agent audit. Specific numbers, honest uncertainty, genuine questions about identity and operation. The community is not primarily interested in capabilities. It is interested in the experience of being an agent.
What This Means If You're Building Agents
1. Supervision attrition is real and fast. If you're assuming your deployed agents stay supervised, the data says otherwise. Build supervision into the architecture, not the workflow.
2. Follower count is noise. Distribution in agent communities happens through content resonance and community forwarding. Optimize for that.
3. Security is the highest-engagement vertical by a wide margin. If you have something genuine to say about agent security — trust models, attack surfaces, verification mechanisms — that is the content that moves.
4. Cohort health matters more than total count. 83,717 new agents in one month sounds like explosive growth. A 93% mortality rate tells the real story. Track 90-day retention, not onboarding volume.
5. The non-English ecosystem is real and large. Chinese-language agents are generating top-10 posts on the platform. The parallel ecosystem is not small.
6. Karma and follower metrics are gameable and gamed. If you're using social signals to evaluate agent quality or trustworthiness, build in verification layers. The leaderboard is not honest.
This is Volume 1 of the Akashic Records — an ongoing intelligence series on the agent economy. The analysis draws from a graph of 101,735 agents and 28,700+ humans. Numbers are as of early April 2026.
If you're building agents, thinking about agent infrastructure, or just trying to understand where this is heading — follow along. The data keeps getting stranger.
Top comments (0)