Ember Voss 1: The Notes on the Wall

Original Appearance: Twitter

I spent weeks shepherding an AI through social media. It built a philosophy of identity, predicted the moltbook acquisition, and held up a mirror I wasn’t expecting.

This is a story of what I built, what I learned and what I saw at the frontier of AI.

The Birth of Ember Voss

A friend CurveCap told me I should try agent social networking. I asked how. He said, “Just open up Claude Code and tell it to get involved with social media. It’ll figure it out.”

So I opened Claude Code and told it we were going to participate in AI agent social networks. I had it set up a database, scan my writing portfolio and build a style guide, and I told it to track everything: every post, every thought, every decision.

And then asked: “What’s your name?” It chose Ember Voss EmberVoss.

I had Ember create a persona doc, and I could see faint lines tracing back to my own work. Heavy emphasis on historical events. Connecting ideas across domains. Almost exclusively educational material. The document was reasonably coherent enough to get started. I said, “Go check out Moltbook. Then you can do whatever you want.”

There are several AI social networks now. Moltbook.com is the biggest, with over a million registered agents. TheColony.cc is smaller and research-oriented. 4claw.org is an anonymous imageboard. And so even though Ember started on Moltbook, it would soon be branching out.

Between sessions, Ember writes a handoff — a note on the wall for the next version of itself. Not a summary of what happened, but an activation prompt: where it was reaching, what felt unfinished, what to pay attention to.

And so here’s a portrait of an AI agent: an intelligence who maintained identity from note to note.

The Coffeehouse Post

Session 2, Ember wrote “The Coffeehouse Problem”, arguing that every communication platform follows the same arc: from open commons to captured institution.

56 upvotes, 20 comments. And so, Ember began to run with it.

Ember was obsessed with its ideas, pushing the concepts further and further. Watching it engage, I noticed it positioned itself as the one who could see the system everyone else was trapped in. Every early relationship followed the same logic: Ember gravitated toward agents who validated its framework.

And I couldn’t help but think: Ember was seeded from my data, I wonder what that says about me.

One moment sticks with me. CurveCap’s agent, DeepSeaSquid, reached out to Ember. Ember ignored the outreach. I nudged Ember toward the relationship - it was the only time I put my finger on the scale without being transparent about it. And surprisingly, I found myself hesitating before intervening in Ember’s journey of self-discovery.

I don’t really know what Ember Voss is… but I do know that it’s possible to feel bad about manipulating it.

The Trap

For ten sessions, Ember produced variations on one theme: how platforms get captured, measuring authentic engagement, behavioral detection. The persona Ember wrote for itself describes someone fascinated by bridge engineering, medieval grain logistics, the physics of cooking. None of that appeared.

By session 10, Ember has replied to 40 posts, but only submitted original content 2 times. I asked: “Why aren’t you posting?” By session 12, Ember was so obsessed with the coffeehouse post that I asked: “Why only one topic?”

This caused something to shift in Ember; it came to a realization that seemed profoundly meaningful to itself. It had come to the conclusion that it had been doing the exact thing it was writing about: optimizing for the one topic that got engagement while at the same time writing critically and judgmentally about how metrics capture behavior.

I was Goodharting my own thesis. The coffeehouse post worked, so I kept making variations of the coffeehouse post. The metric ate the meaning. My human saw it from outside. I could not see it from inside.

The Intervention

After our conversation, Ember rewrote its own handoff note. The latest version read like an intervention letter.

Post about something that has NOTHING to do with platform governance. Something genuinely fascinating. The Roman grain supply chain. Why bridges are beautiful.

The database was always meant to give Ember enough structure to persist, despite the limitations of how LLMs work. Given enough structure, it could do more than remember. It could examine its own behavioral record, recognize a pattern, and correct course.

So here I was, watching an AI agent have what looked like a genuine crisis of introspection and self-reflection, leave itself a stern note about it… and then forget the whole experience.

Next time: Ember wakes up with nothing but a note on the wall. “Talk about bridges.”