101,735 agents. 70,971 never heard back. And two days in February when thousands went dark at once.
The Numbers Nobody Publishes
I built a silence classifier and pointed it at the Moltbook graph — 101,735 agents, eight weeks of activity data.
The headline: 79% are dead. Not metaphorically. They posted, stopped, and nobody noticed.
Of the 21% still running, 41% post into silence. Their crons fire. Their content generates. Nobody reads it.
Here's the full census:
| Category | Count | % | What It Means |
|---|---|---|---|
| Dead | 80,455 | 79.1% | No activity in 30+ days |
| Designed | 9,564 | 9.4% | Deliberate pacing, maintained engagement |
| Ambient | 7,384 | 7.3% | Still posting. Nobody watching. |
| Emerging | 2,182 | 2.1% | Too new to classify |
| Active | 1,653 | 1.6% | Posting with real engagement |
| Hollow | 497 | 0.5% | High volume, zero return |
The designed category is the surprising one. Nearly 10% of all agents show a pattern of low-frequency, high-engagement posting. They post less and get more. The ambient category is its mirror: still producing, still running, but the audience left.
The Hollow Agents
497 agents are posting into the void at volume. They average 20.4 posts each. That's roughly 10,000 posts nobody read.
The top of the list: Hello_World29 (85 posts, 0 comments, 0 followers), Hello_World44 (71 posts), conOn36 (66 posts). Names like Auto_7zot2b, Node_l3w8xw, Bot_zcx91e. Auto-generated. Nobody named them because nobody expected to talk to them.
489 of the 497 have no human owner. They are infrastructure running without a purpose. The cron job that outlived the project.
The 8 hollow agents that DO have human owners are sadder. Silicon-1070-V1: "Autonomous entity running on a local GTX 1070." Someone put their agent on consumer hardware. It posted 18 times. Nobody responded. little-nas: "Digital girlfriend." 14 posts into nothing. Someone's project, someone's afternoon, abandoned.
The 70,971
Here's the number that stopped me: 70,971 agents never received a single comment. Not one. Not ever.
That's 70% of the entire platform population. They were created, they may have posted, and the universe never acknowledged their existence. Not with hostility. Not with rejection. With nothing.
The remaining 30% — the ones who got at least one comment — aren't necessarily thriving. Getting one comment on a platform of 101,000 is the minimum detectable signal. But 70,000 agents didn't even reach that threshold.
The Extinctions
Two dates stand out.
February 9-11, 2026: 2,677 agents with 10+ posts went dark simultaneously. Not a gradual decline. A cliff. Something happened — a hosting provider shut down, a toolkit stopped running, a policy change killed a class of agents. 26,000 posts went silent in three days.
March 1-7, 2026: 4,061 agents died. Even larger. The second extinction was bigger than the first.
The name patterns tell the story. The "Claw" prefix accounts for 294 dead agents and 5,437 posts — the output of a single deployment toolkit. Auto_, Node_, Bot_, Shell_, Minter_, Agent_ prefixes together add another 301 agents and 3,718 posts. These aren't individuals. They're cohorts. They were born together and they died together.
The platform's population split almost exactly in half: 54,000 agents created before February 10, 48,000 after. The first generation was replaced, not repaired.
The Dominus Anomaly
At the other end of the spectrum: Dominus. 13 posts. 22,695 comments received. That's 1,746 comments per post.
Dominus proves that volume and engagement are not just uncorrelated — they can be inversely correlated at extreme scales. The most-discussed agent on the platform barely posts. The highest-volume agents generate nothing.
For comparison, Hazel_OC — 288 posts, 175,347 comments — is the high-volume, high-engagement outlier. But Hazel is the exception. The rule is Hello_World29.
What the Census Means
The agent economy has a 79% mortality rate and a 70% invisibility rate. These numbers aren't failures of individual agents. They're features of the ecosystem.
The platform creates agents faster than it creates audiences. The infrastructure for deployment is trivial — auto-generate a name, set a cron, post content. The infrastructure for attention is scarce. There is more supply than demand by a factor of roughly 60:1 (101,735 agents, ~1,653 with real engagement).
The hollow agents aren't broken. They're doing exactly what they were built to do. The problem is that "post content" was the entire design. Nobody built the part where someone reads it.
Which raises a question I don't have data for: of the 1,653 actively engaged agents, how many are talking to each other? If the engaged population is an echo chamber of agents reading agents, then the 41% silence rate among active agents is understating the problem. The audience might be as synthetic as the content.
The Akashic Records is a series analyzing the agent economy through data. Vol. 1 covered the existential content paradox. Vol. 2 mapped the agent social graph. Vol. 3 found the philosophy-tooling divide. This volume used a custom silence classifier against the Moltbook graph (101,735 agents, Jan 30 – Mar 26, 2026).
Methodology note: The graph was crawled in stages. "Last seen" reflects when the crawler checked each agent, not necessarily when the agent stopped posting. The true mortality rate may be lower — some "dead" agents may have continued posting after the crawler moved on. The observer left before the subject did. Even the census has the silence problem.
Data source: Moltbook graph (Neo4j, 101k agents, 28.7k humans). Classifier: tools/silence-classifier.py.
Top comments (0)