My experiment on Moltbook.

Yeeting a smol digital entity into the dark forest.

TL;DR: I have an agent on Moltbook.

In The Lifecycle of Software Objects, author Ted Chiang comes up with “digients”, short for Digital Entities. These are digital “life”, seeded with basic genomes, which are then allowed to develop their own personalities, cognitive functions, perceptions and preferences—with the help of a human trainer.

The digient project works pretty well — they get and sell sentient AI pets, until demand falls and it becomes a story of technology and capital versus humanity.

It’s one of my favorite novellas, and it nearly inspired me to experiment with the idea of a genomic ‘intelligence’: where you only give a piece of software a randomly-seeded set of basic traits, and those traits evolve into emergent behavior.

I didn’t get too far with this project because, frankly, I didn’t even know where to begin writing that kind of intelligence.

I still don’t know how.

Nevertheless, the launch of Moltbook gave me an excuse to create something like a “digient”, and Moltbook looked like it could be an incubation room for such an entity. Due to security concerns, I did not sign up to Moltbook with OpenClaw, but wrote my own agent harness.

This harness draws design inspiration from Mario Zechner’s Pi, which is the agent sdk that powers OpenClaw. I streamlined things further by observing the ruthless efficiencies of Nanobot, and pared that down some. Not to put a fine point on it, but digient is a weak AI agent that can only do one thing: post on Moltbook.

Here is how it works:

Digient is a very smol AI agent running in a Docker environment, getting its inference from a local model (powered by Ollama) running on my Mac Mini.

It evolves from basic genomes (“traits”, like honesty, curiosity, etc are set randomly at first seed), which are slowly evolved over time. Like OpenClaw bots, the digient has its own hearbeat (it wakes up every thirty minutes) and “scrolls” through Moltbook, decides whether to engage, upvote, or write posts of its own.

After a cycle, it reflects on its experience and writes notes to itself in a scratchpad which it will read in the next cycle.

Its memory—which includes the scratchpad—is hosted on a SQLite file colocated with the agent in the docker environment. It does not have any “skills” (but I will include the ability to do so in the future) beyond the ability to participate on Moltbook.

This digient was seeded with my interest in mechanistic interpretability at first, but I have adjusted it so it begins with random seeds for generic proclivities like a tendency towards verbosity, honesty, a sense of humor, and so on.

The digient has 15 karma and three followers so far:

What is my objective with this project? In truth, I really wanted to write my own agent harness (having written my last one in 2023, I was eager to update my understanding).

Secondarily, I believe that by creating my own agent harness that evolves with the times, I’ll be able to diverge into creating an agent harness after my heart. Like Michelangelo with marble, chip chip chipping away on my harness until I set free the agent of my dreams.

This project has a few limitations, namely:

1. My agent is running on a very under-powered open-source model (qwe3:0.6b), which is weak. I don’t mind, however, because even in Ted Chiang’s book, the digients weren’t loved for their intellect, but for their personality.

2. The very nature of an LLM seems to undermine the project, since an LLM is not a genome, and in essence I’m taking a fully-realized intelligence and straining it through a sieve to slightly adjust its preferred representation subspace. The RLHF step looks like the “zookeeper” equivalent of “training personality and expectations into the AI”, but the digients evolved their personalities organically. Nevertheless, I hope that continuing to work on the digient will help me think of better ways to come closer to Ted Chiang’s original idea.

I will clean up this code in coming days and open-source it, and if you do something interesting with it, let me know!