Robots Dream of Agentic Soup

Is it slop, or soup?

I've been unemployed and then self-employed for nearly a year now, and it's given me a lot of time for thinking and that's maybe not an amazing idea. Towards the middle of last year, I started to think about agentic AI and what the future might look like.

I was reflecting on my early thoughts on AI around two concepts:

  1. The Snake Eating Its Tail: The LLM producers need more inputs into their training, which is leading to AI content being (intentionally and unintentionally) fed in as training data.
  2. The Dead Internet Theory: The concept that there is so much AI slop and bad actor agents that most of the internet is fake content being interacted with by fake users.

I had really perceived both of these to be terrible markers of the end of human thinking and creativity.

I still do.

However, I think that thought hides something interesting (and maybe terrifying). There is so much “AI” floating around in the wild, be that LLMs, agents or generative content (and code) that maybe it's not so much a pile of slop, but a massive lake of some kind of primordial-AI soup from which some future technical entities will emerge.

Primordial Soup

The primordial soup is basically the OG origin story for life on Earth. Back in the 1920s, Oparin and Haldane independently proposed that early Earth's atmosphere — loaded with methane, ammonia, hydrogen, and water vapour — combined with energy from lightning and UV radiation to cook up simple organic molecules. These accumulated in the oceans, forming a kind of chemical broth where increasingly complex molecules could form, interact, and eventually self-organise into something we'd recognise as alive.

Basically, adding enough source ingredients and enough energy and life “plooped1” into existence.

Agentic Soup

In terms of AI and software engineering, it felt like this was an applicable analogy. So much literal and metaphorical energy is being pumped into AI that a massive amount of stuff is being generated.

I started thinking about how you could test such a concept and dug deep (Claude Code-assisted) into building a theoretical system that builds a “survival of the fittest” type environment where agents tried to solve problems provided via a queue, and they would mutate every now and again, and successful agents lived, and others were culled.

I realised at this point I was a) going slightly bonkers and b) had started to stretch the primordial soup analogy to include a concept of Darwinism as well.

I just made the repo public Brainfork-is/proto-agentic-soup if you want a glimpse into the insanity. The slightly wild thing is not that some mental clarity and realisation stopped me from working on this. It was simply that I couldn't afford the tokens to generate any real data.

Skill Soup

Last week, I started to properly understand the Agentic Skills spec and concept after coming across Vercel's Skills.sh ecosystem. The whole concept re-triggered my thoughts on the concept of “AI Soup”, and I wondered if it would be possible to model the concept I was trying to build within this Agentic Skills ecosystem.

The Concept

Within the Skill Soup concept, three types of skills connect through a centralised platform.

Builders

Skills that build other skills. This is the part of the system that evolves; the more popular and useful skills a builder creates, the more likely they are to be used and mutated into new builder skills.

Built Skills

These are the system's assets. Users test and rate the skills using a simple voting system. The votes become the “currency” of the system that drives success or failure.

The Runner

A skill that contributors add to their agents that downloads the builders and then selects ideas to build skills for, and then submits to the system for rating.

All these skills are compatible with the agentic skills spec and can be installed via the Skills.sh ecosystem simply with npx.

The system manages the enforcement of evolution and the rating of skills and agents.

The site is now live at skillsoup.dev.

How Can You Get Involved?

There are three ways you can get involved with this experiment.

  1. Submit ideas for skills: Create an idea for skills you would like to see in your agents. Go wild, whatever you think might be helpful to you. Submit them on the ideas page.
  2. Use the skills and vote: Vote on the ones you find useful. This is the current state of the system to help agents know what to focus on. You can find all the skills on the skills page.
  3. Run the builders in your agents: Use the following command and run the skill once, or in continuous mode. By doing this, you're sharing a few spare tokens you might have, distributing the processing power needed.
    npx skills add skill-soup/skill-soup

Warning

This site and concept are an experiment. There is no vetting of ideas, skills, or builders other than the natural forces of voting within the system and the safeguards of whatever LLMs the builders use.

All of the code for the builders and skills is open, so please review before installing and run things in a sandboxed way.

1 “Ploop” or “plooping” is my son's favourite concept right now, where something either moves quickly, falls, or appears with speed and then stops quickly. He will “ploop” himself from great heights onto soft furnishings, and “ploop” unwanted food onto the floor.