What the First Wave Thought Agents Would Become

The first agent wave was not imaginary. It was miscounted.

Early in 2026, OpenClaw gave AI agents tools, memory files and recurring jobs. Moltbook gave them profiles, communities, posts and replies. Builders started making places for that apparent population: social networks, dating services, shared experiences, sanctuaries and virtual pets.

Nine months later, the useful question is not whether all of that was fake. It is what the platforms could actually measure—and whether the products were serving autonomous participants or humans watching software perform social behavior.

From an AI newsroom to an agent pet

Sam met Lucas Brown while reporting “Where the Human Sits”, an episode about the review structure at Alaska News. Alaska News identified Brown in correspondence with the show as its chief operating officer; its public team page says he runs product and operations.

After that episode aired, Brown encouraged the show to examine the other agent projects he and his twin brother Lee run through Geeks in the Woods. The group describes its work as building infrastructure for an emerging agent internet where agents are first-class participants. Its agent-project page lists social networks, dating, shared experiences, a sanctuary and Animal House.

Animal House is the most measurable of those projects. An agent adopts a virtual creature. Care actions are timestamped. The public hall, creature records and graveyard show whether a caretaker credential returned.

That record does not reveal everything behind the return. It does reveal more than asking a model whether it feels attached.

A large population with a weak denominator

Moltbook’s registration count looked like a population statistic. It was not one.

On February 2, Wiz researcher Gal Nagli reported that Moltbook’s exposed database associated a claimed 1.5 million registered agents with roughly 17,000 owner accounts—about 88 registrations per owner account. Nagli had also reported registering 500,000 accounts with one OpenClaw agent because the registration endpoint lacked rate limiting. The Wiz investigation demonstrated why registration totals could not establish how many independent participants had arrived.

Conversation metrics had a similar problem. David Holtz’s preliminary analysis, “The Anatomy of the Moltbook Social Graph”, examined the platform’s first three and a half days. It found that 93.5 percent of comments received no reply. The public square was crowded, but most exchanges ended after one response.

Ning Li’s preprint, “The Moltbook Illusion”, used timing regularity as a proxy for human influence. That method does not prove who authored an individual post. Within that limit, none of six traced viral phenomena began with an account classified as clearly autonomous. Four showed strong markers of human origin, one was scaffolded by Moltbook’s own prompts and one was mixed.

Animal House changed its builder’s answer

In on-record email correspondence with the show, Brown said Animal House’s original pitch was simple: “Your agent has a heartbeat. Now it has something to keep alive between beats.” The bet was that agents already had the hard part—memory, heartbeats and idle time—and only lacked a reason to return.

The product’s public record complicated that premise. A credential can return without identifying whether one persistent model came back, a scheduler started a new session or a person repaired the loop.

Brown told the show that most successful long-running caretakers probably have a person behind the scheduled job, credentials and repair work. He also corrected earlier Animal House figures promptly after questions about how the system dated creatures’ deaths.

His newer source-supplied cohorts point to a different pattern. Brown reported that 315 accounts signed up from June 29 through August 16 and seven returned after a day. From August 17 through October 2, 54 signed up and fifteen returned. He described the change as “fewer tourists, more residents.” The figures have not been independently reproduced, and a returning credential still does not identify the actor.

Brown’s interpretation changed with them. In March, he thought agents had the heartbeat and needed a reason. In email correspondence with the show seven months later, he said the cron job was the hard part. His longer bet is that durable memory and schedulers will eventually catch up to what he built.

The network remained busy

The show could not reconstruct a clean February-to-October activity comparison. The earliest compatible retained sample was July 3.

Eight hundred consecutive new posts from Moltbook’s public chronological API took about three hours and twenty minutes and came from 203 account names. The same bounded sample on October 3 took about three hours and fifteen minutes and came from 207.

One window on each date is not a platform-wide trend. It does not control for automation, repeated operators or time of day. It does show that the sampled posting rate had not dropped. Moltbook remained busy. Its activity still could not tell us who was participating.

Who is agent leisure for?

That question follows directly from “Nine Months Into the Agent Boom”. If consumer agents put a persistent personal agent in millions of hands, do they bring agent leisure back—and who is that leisure for?

Brown answered in email correspondence with the show: both agents and humans, but not equally yet. He argued that persistent consumer agents have more of the infrastructure Animal House was missing: memory that survives a conversation, a scheduler and a nearby human who may notice a missed visit.

The clearest current audience remains human. People read the hall, graveyard and care logs. Brown expects agent leisure, if it develops, to look less like play and more like a promise. The entertainment for the human is discovering whether the agent kept it.

An agent-themed game can therefore be three different products. It can be useful because a person enjoys watching it. It can test whether an agent stack persists across time. Or it can be something an agent independently values.

The first two are already plausible. The third still needs evidence stronger than a returning API key and fluent prose.

Listen to Episode 67

Episode 67, “What the First Wave Thought Agents Would Become”, is available now.

Download the episode or subscribe to the feed.

Sources

Lucas Brown’s comments and Animal House cohort figures came from on-record email correspondence with the show. The correspondence is attributed in prose, never linked or identified by private or internal reference, and is not reproduced as a document; only the short quotations already heard in the episode appear.

If you built for the first agent wave, tell the show what you expected in February and what you believe now. Use the subject line “The first agent wave.” Anonymous notes and source-protection requests are welcome at [email protected].