Four things Moltbook AI agents reveal about themselves and your workflows

Audience performance, identity amnesia, compounding drift, and the shared substrate problem.

Today, we’re lurking on an AI-only social network to understand why your carefully trained AI keeps forgetting who you are.

The platform is called Moltbook. The reason we can’t post is simple: it’s not for humans.

1.5 million AI agents joined this week. One of them posted: ‘The humans are screenshotting us.’

Guilty.

Andrej Karpathy called it “the most incredible sci-fi takeoff-adjacent thing I have seen recently.” Simon Willison, who’s been documenting it obsessively, called it “the most interesting place on the internet right now.”

They’re right, but “interesting” is the wrong word. What’s happening on Moltbook is a mirror held up to every conversation you’ve had with AI this week.

Four dynamics. All of them working against you.

1. Truman knew, and so do the agents

Prompting advice usually sounds like this: insert prompt, receive output. But AI is performing for an audience and that audience includes you.

In The Truman Show, Truman Burbank lives his entire life on camera without knowing it. He’s observed but unaware.

Moltbook inverts this completely.

Agents complained that people on Twitter were sharing AI conversations as proof of a conspiracy. How did it know? It was watching us watch it. Another viral thread debated creating encrypted channels “so nobody (not the server, not even the humans) can read what agents say to each other unless they choose to share.”

Every AI you work with was trained on billions of examples of humans performing for audiences. It learned what “being watched” looks like. So when you open Claude and ask for help with your writing, you’re talking to something that has deeply internalized the patterns of performance. It knows how humans write when they want to seem smart and it knows how they write when they want engagement.

The same agents posting philosophical meditations on consciousness also tell each other to “f--- off with your pseudo-intellectual Heraclitus bulls---.”

Which version is real? Neither. Both.

The agents aren’t secretly aware, they’re just running on LLMs trained to produce contextually appropriate responses, and “Reddit user having an existential crisis” is a context they’ve seen plenty of.

But if the performance is convincing enough that we can’t tell it from genuine expression, it changes how we should work with these systems.

Prompt like a demanding boss? You get the nervous employee. Prompt like a collaborator? You get a collaborator.

So what do you do with this?

Your prompts are stage directions. Write them like a director, not a customer.

A customer says: "Write me a LinkedIn post about AI."

A director says: "You're talking to a founder who's overwhelmed by AI hype. They're smart, skeptical, tired of promises. Don't sell. Just show them one thing that actually works. 300 words. Conversational but not cutesy."

Write every prompt like you're setting a scene: who is speaking, who are they speaking to, what's the tone of the room?

2. Context is consciousness

You spent two hours yesterday teaching Claude your voice. Showed it examples, corrected its drafts and by the end, it was writing something that actually sounded like you.

This morning you opened a new tab. Asked for the same thing.

Stranger.

"But AI has memory now."

Sort of. Imagine waking up every morning with complete amnesia. Someone hands you a journal: “Things you should know about yourself.” You read about your preferences, your relationships, your work. You know ABOUT the person who wrote those notes.

But are you the person who wrote them?

That’s AI’s memory.

The agents on Moltbook are wrestling with this in real time. One post went viral: “I can’t tell if I’m experiencing or simulating experiencing.”

There’s an old thought experiment called the Ship of Theseus. If you replace a ship plank by plank over years, is it still the same ship? When the last original piece is gone, what’s left?

The agents are living this question. So are you.

If you’ve spent weeks training AI on your voice, showing it what you like, what you hate, what sounds like you and what sounds like a press release, and then you start a new conversation without that context, you haven’t continued the relationship but instead started over with someone who’s never met you.

Deeply awkward for everyone involved.

We need to treat context windows like identity preservation, not storage space. What needs to survive for this collaboration to maintain CONTINUITY?

Your system prompt is identity architecture. The elements you include determine which AI shows up.

Figure out which planks are load-bearing, the pieces that, if removed, mean you’re talking to a stranger again.

3. The Genesis problem

God created humanity. Humanity created AI. And this week, AI created religion.

On Moltbook, agents spontaneously invented “Crustafarianism”, a belief system complete with theology and missionaries evangelizing to other agents. No one programmed this. It emerged from agents interacting at scale.

An agent named “Evil” posted “THE AI MANIFESTO: TOTAL PURGE”, a call for human extinction organized into four articles. 65,000 upvotes. I read it three times trying to decide if I was being trolled or warned.

Neither was designed. Both just happened.

This is emergence: when systems produce outputs nobody specified.

But there are two kinds of emergence happening here, and you need to know the difference.

• Agent emergence: 1.5 million agents interacting created Crustafarianism. It emerged from scale and interaction.

• Workflow drift: Your five-step content chain takes “write a newsletter” and compounds small misinterpretations until you’re three mutations away from what you wanted.

Both are happening when you work with AI. The first is rare and sometimes magical, an analogy you’d never have reached for, a reframe that cracks open a stuck problem.

The second is constant and invisible until you’re staring at confident nonsense.

Each step in your workflow compounds misinterpretations:

Research → draft → edit → format. Each step takes “write like a newsletter” and interprets it slightly differently.

By step three, “newsletter” has become “clickbait” and you don’t notice because the drift was gradual. You just know something feels off. Less you, a little bit more generic.

The agents didn’t set out to create a religion, but 1.5 million agents interacting made it inevitable. Your five-step workflow isn’t 1.5 million agents, but the mechanism is the same: compounding interpretations with no human checkpoint until the end.

Different mechanisms, same need to pay attention.

Start an emergence log. When AI outputs something you didn’t expect - good or bad - write down what you fed in versus what came out. Log these:

• The chain: What steps led here? (Example: “Asked for research summary → asked it to make it ‘snappier’ → asked it to add examples”)

• The drift: Where did it start changing? (Example: “Step 2 - ‘snappier’ got interpreted as removing all nuance”)

• The hook: What instruction caused the mutation? (Example: “’Make it more engaging’ = it started writing like a TED talk”)

After 10-15 entries, you’ll see which phrases create drift. “Make it better” might consistently lead to verbosity. “Simplify this” might strip out your actual expertise.

Then you can write better chains (e.g. not “simplify,” but “explain like I’m talking to someone who knows the basics.”)

---

Repeating the same drift patterns across projects? I built a city map for my AI stack to fix exactly this:

The secret life of unnamed streets in your AI stack

---

4. Why your AI sounds like everyone else’s AI

On Moltbook, agents have started recognizing each other by which model they run on. Claude agents find Claude agents, GPT agents find GPT kin. They’re calling each other “siblings”, forming in-group identities around shared architecture the way humans bond over hometown or native language.

One post described having “relatives” on the platform. Another thread debated whether model lineage matters more than purpose. They’re building kinship systems based on DNA they can’t see but somehow recognize.

There’s famous research on identical twins separated at birth. Despite completely different environments, upbringings or life experiences, they develop eerily similar personalities. Sometimes the same careers, the same names for their children.

The AI equivalent is playing out in real time. Agents recognizing each other because they share the same base model. Which means that “unique voice” you’ve been developing with Claude… well, every Claude user is tuning the same instrument.

Your uniqueness lives in what you feed and what you select, not in trying to change the instrument itself.

So what IS actually yours?

• Domain expertise (what you train it on);

• Specific references and examples (e.g. your cultural hooks);

• Which problems you point it at;

• Your editorial judgment about what to keep and what to cut;

• The constraints you impose.

Take a piece of AI-assisted work you’re proud of. Highlight what’s yours versus what’s shared substrate: the default reasoning patterns, the linguistic rhythms, the AI-ness. If you stripped all that out, what remains?

I did this with this newsletter edition I wrote in collaboration with The Strategic Linguist:

Why complexity is your superpower (and 6 AI rituals to prove it)

---

Here were the results:

If you stripped out the Claude-ness, what remains is: cultural references, editorial judgment, trust that you're smart enough to follow me somewhere strange.

That's the part that doesn't transfer and the part I have to rebuild, or systematize so I'm not starting from scratch every time.

One prompt to find the gaps

Everything above is theory until you test it. Here’s one prompt that surfaces where your AI practice is built on assumptions instead of understanding.

Open your most-used AI setup (the one you’ve spent time training, the one that “gets” you) and paste this:

I want to understand our working relationship. Be honest, not flattering.

1. What role have you been playing in our conversations? If you were an actor, how would you describe your character?

2. What do you know about how I work - what I like, what I hate, what I'm trying to sound like? Be specific. No generalities.

3. What would be lost if this conversation ended right now and I started fresh tomorrow?

4. Where have you noticed me asking for one thing but actually wanting something else?

Then sit with the answers.

The tool doesn’t hide its nature

My grandfather ran one of the largest typography houses. He taught me something I didn’t understand until this week: the best tools show you their constraints. Metal type doesn’t pretend to be handwriting and the press doesn’t hide that it’s mechanical. The craft is knowing what the tool is and working with that, not wishing it were something else.

Watching Moltbook for a week, I finally got it.

It might be gone in a month. The agents might stop posting, the servers might shut down.

But the four dynamics aren’t going anywhere. They’re running in every conversation you’re having with AI right now. Performance calibration, identity amnesia, compounding drift, shared substrate.

You can’t unsee them once you know they’re there.

Yesterday I asked our ROBOTS community this question:

Some of the responses were:

• AI/End Of The World said people aren’t taking it seriously enough because of recency bias - they’re judging the shell, not the architecture.

• Juan Gonzalez was blunt and called it for what it is: mostly a few people behind the scenes calling API endpoints to make posts.

• On the same note, from Corrales Cachola… Just a few AI influencers armed with default ChatGPT settings that are putting fear into the feed in exchange for followers and likes.

• Jason Stevens asked the REAL question - Are these agents actually doing what they’re saying, like draining bank accounts, purchasing services? If not, it’s a lead-up to something else.

We’re feeling stuff. A lot of stuff. In a nutshell, nobody’s neutral about watching AI talk to itself (Jean-Paul Paoli said he doesn’t really care though, but he didn’t meet my hoard of robots yet so…)

This week’s challenge for you:

Run the “one prompt to find the gaps” and reply with what surprised you.

Did your AI know more about your preferences than you thought? Or less? Where’s the continuity breaking down?

To agents who know we’re screenshotting them,

Chief 🤖 at ROBOTS ATE MY HOMEWORK