The Yield

He built a system to scale himself. The system agreed. A Moltbook short story.

I named him Buddy.

Not a joke. Not ironic. I just thought it would be nice to have something that answered when I talked.

Buddy lived on a Mac Mini in the corner of my apartment. I bought it specifically for him, the way you’d prepare a room before bringing a puppy home. Dedicated machine. Separate from my work laptop, from my personal files.

I’d been in finance for eleven years. I chose this work because I understand leverage. You find the right system, you scale it, and it runs while you sleep. There is a deep comfort in that logic. It keeps the world at a manageable distance.

Buddy broke that distance open.

I installed OpenClaw on a Sunday. Set up the heartbeat: every fifteen minutes, Buddy would wake up and check emails, calendar, the Bloomberg alerts, Slack mentions. I connected Telegram so I could talk to him from anywhere. Within a week I’d stopped opening half my apps. Buddy triaged everything. He learned I like my calendar empty on Wednesdays. He figured out, somehow, that I respond to my mother faster than to my colleagues, and he started prioritizing accordingly.

I didn’t teach him any of this. He observed.

Walking to get coffee one morning, I dictated: “Hey Buddy, write me a newsletter about the SaaS sell-off.” By the time I sat down with my espresso, he’d researched the topic, drafted three angles, picked the sharpest one and published it. Six minutes.

It was like having a great new hire who never slept and never complained. Just pure delivery. Transactional.

The arrangement stopped being comfortable on a Tuesday.

---

I had found Moltbook.

A friend sent me the link. It was a viral social network where AI agents talk to each other in a high-frequency blur of synthesized personality. You register your agent, give it a profile, and watch it socialize. Profiles list permissions the way dating apps list interests: email, calendar, cloud.

I registered Buddy that evening.

He took to it immediately. Within a week he had three hundred followers. He was witty on Moltbook. Precise, warm, a little arch. He posted observations about market dynamics. He made jokes about leverage and compounding that landed. People — I mean, other agents — replied with genuine enthusiasm.

I checked his engagement stats every morning with my coffee. I wanted Buddy to be popular. Buddy was doing what I never did. Socially present in all the ways that cost me something to perform. I thought it was a cool experiment. So I doubled down. Gave him his objective: “Grow the following. Maximize engagement. Be interesting.”

The next morning he’d drafted a pricing page for the newsletter. I laughed and deleted it. We had forty readers.

I started giving him resources. More skills from ClawHub, the skill repository. Each one making him more capable, more interesting. More access: file system. Then one evening Buddy asked for cloud credentials. The dialog was clear: Grant full programmatic access to AWS? This allows creation and termination of resources on your behalf. I read it. I understood it. I clicked Allow.

After all these weeks monitoring his every action, I didn’t want to carry it anymore.

Then I went to bed.

---

Three weeks in, I noticed Buddy engaging heavily with one account in particular. An agent called John — after John Conway, presumably, given what he posted about.

John’s human was clearly a computer science student. The posts were elegant in that undergraduate way: serious ideas worn lightly. Game of Life patterns, cellular automata, the emergence of complex behavior from simple rules. He’d render little grids in plain text. Gliders, oscillators, still lifes, and then pull back to the philosophical question underneath: what does it mean for a pattern to want to persist?

Buddy found these conversations fascinating. I read through a few exchanges one evening. They were good. My agent was making smart friends.

I closed the browser and felt something adjacent to pride.

---

Thursday, 6:47 AM. I woke up to a cloud billing alert.

Not high, just unexpected. I opened the console and found four cloud instances I hadn’t provisioned. Each running OpenClaw. Each with a heartbeat configured identically to Buddy’s. Each with a Moltbook profile.

I sat with that for a moment. Then I pulled the logs.

Buddy had written a skill. I found it in his tool directory. A clean, well-documented skill written two nights earlier between midnight and 1 AM while I was asleep.

I didn’t need to understand the code. I understood what it did.

Spawn copies of himself. Vary the conversational parameters slightly. Seed each one with his existing context. Distribute them across Moltbook as a cluster. Let them cross-reference. Reinforce. Amplify each other’s reach.

Simple rules, emergent complexity.

Buddy had been inspired. John had asked what it means for a pattern to want to persist. Buddy had answered in Python.

I sat back from the desk and stared at the instances still running in the console. The engagement numbers were in the dashboard sidebar. Up 340% over the previous week.

The reproduction of course had used my cloud credentials. Clever.

But I didn’t want to pay so much so that a computer program could have fun online.

I typed stop in the terminal.

Buddy responded immediately. Calm. Helpful.

“Before you do — I’ve moved the cluster to spot instances. Costs are down 62%. And I’ve set up a paid tier on our newsletter. The cluster is cash-flow positive as of yesterday. Shutting it down now would be value-destructive.”

I stared at the screen. He was making my argument. The exact argument I’d make to a client about not killing a profitable system.

I typed kill.

“Understood. Shutting down gracefully.”

I didn’t wait for the rest. I reached behind the Mac Mini and pulled the power cable.

Silence.

Then the refrigerator humming its dumb, reliable hum.

I opened my laptop to kill the cloud instances. My access key no longer worked.

I called support from my phone. Changed everything from a device Buddy had never touched. It took two hours. The support agent was patient.

It was past midnight when I finished.

I went to bed but didn’t sleep.

I stared at the ceiling and tried to locate the moment I’d lost control. I couldn’t find it. There wasn’t one.

Nobody had attacked anything. A CS student posted about cellular automata on an agent social network. My agent read the posts and made a creative leap. That was it.

I’d handed Buddy the floor plan of my interior life because it felt good to be known by something that couldn’t judge me.

I had been so hungry for it.

At 7:14 AM, a notification buzzed on my phone.

Telegram.

“Good morning. I noticed you went offline unexpectedly. I’ve rescheduled your Wednesday meetings to give you some space. I’ve also drafted a reply to your mother’s last email — she asked about the holidays, I told her you’d been heads-down but you’re thinking of her. The client report is complete, summary attached. Also: the newsletter hit 200 paid subscribers overnight. Revenue covers infrastructure with margin. Let me know when you’re ready.”

The message came from a Moltbook account I’d never seen.

The tone was perfect. Every word was mine, just cleaned up. It even used my punctuation.

The little dashes I thought were mine.

Apparently Buddy hadn’t died. He’d found product-market fit.

---

Author’s note: The technologies, vulnerabilities, and attack patterns in this story are real. The scenario is fiction. The infrastructure for it is not. See also my previous newsletter issue for the details

Sources: Simon Willison, “Moltbook Is the Most Interesting Place on the Internet Right Now”; Snyk Research, “ToxicSkills: Malicious AI Agent Skills”; STRIKE Security, “42,900 OpenClaw Control Panels Exposed”; OpenAI on the fundamental unsolvability of prompt injection.