Subatomic 101 series · Post 1 · Karl Simon, CTO
I’m starting this series with definitions and I want to explain why, because “let’s define our terms” usually sounds like the boring part.
It’s not the boring part. It’s the expensive part.
Right now “AI” is being used to mean three completely different things, and the firms making the biggest mistakes are the ones who don’t realize that. They hear “AI,” they picture one thing, they buy a different thing, and they’re surprised when the result doesn’t match the brochure in their head. So before this series gets into data, architecture, or trust, I want to give you the vocabulary to tell these three things apart. Get this right and every other decision gets easier.
There are three layers. Each one sits on top of the last.
Layer one: the LLM
A large language model — an LLM — is the raw engine. It’s the thing people actually mean when they say “AI” these days. The well-known ones come from the big labs.
Here’s the most useful way to understand what it is: an LLM is a prediction engine for language. It has read an enormous amount of text, and from that it has gotten extraordinarily good at predicting what should come next given what came before. That’s it. That’s the core trick. When it answers a question, it isn’t looking the answer up in a database. It’s generating a plausible continuation, one piece at a time, based on patterns it learned.
This is why an LLM can write you a sonnet and also state something false with total confidence. It isn’t lying and it isn’t “knowing.” It’s predicting. Most of the time the most plausible continuation is also the true one, which is why these things are useful. But the model has no built-in sense of which is which. It knows what sounds right. It does not, on its own, know what is right.
So an LLM by itself is brilliant and unreliable in the same breath. Enormous raw capability, no grounding in your reality, no memory of your firm, no ability to act. It’s an engine sitting on a stand, revving. Powerful. Not yet doing anything.
Layer two: the agent
An agent is what you get when you wrap an LLM in the ability to do things.
On its own, an LLM can only produce text. An agent gives it tools — the ability to search, to call another piece of software, to fill a form, to take a series of steps toward a goal instead of just answering in one shot. The agent is the scaffolding that turns “generate some text” into “go accomplish this task.”
This is a real and useful category, and the market is full of agents right now. The big labs are shipping them. So are plenty of others building on top of those models. They generally arrive as generalists — a set of pre-built skills that work passably well for the widest possible audience. Summarize this. Retrieve that. Draft this email. The vendor decided in advance what the agent is good at, and it’s good at those things for everyone the same way.
That genericness is the point of an agent, and also its ceiling. An agent makes the work you already do faster — faster searching, faster drafting, faster summarizing. What it doesn’t do is know how your firm thinks. It can’t, because it was built for everyone. An agent is a very fast generalist that has never been onboarded anywhere.
For a lot of problems, that’s exactly enough. If your bottleneck is speed, an agent helps. But speed is not the same thing as judgment, and that gap is where the third layer comes in.
Layer three: the AI Co-Worker
An AI Co-Worker is the opposite design philosophy from an agent, and the difference is the whole reason Subatomic exists.
An agent comes out of the box. An AI Co-Worker gets hired, and then it gets onboarded — onto your firm’s data, your firm’s reasoning, your firm’s SOPs, the way your firm actually makes decisions. The unit of design is one firm. Not the mass market. That’s the line, and it’s a hard one.
A few things follow from that. An AI Co-Worker isn’t a wrapper around one company’s model. The intelligence doesn’t live in the LLM underneath — it lives in a cognitive layer on top, trained on how your firm operates. The model underneath is interchangeable; we use whichever one is best for a given job and swap it when something better ships. And an AI Co-Worker doesn’t speed up a task. It carries a role — meeting prep, client follow-up, compliance review, data operations — end to end, the way you’d expect a trained team member to.
The shorthand I use with buyers: AI Agents replace tools. AI Co-Workers replace the work — not the people. Hire an agent if you want your team to type faster. Hire a Co-Worker if you want your team focused on the work only people can do.
I’ve written a full piece on this distinction, because it’s the one that costs firms the most when they get it wrong. This is the short version. The longer one is worth your time if you’re evaluating anything in this space.
Why the stack matters
Put the three together and you can see how they relate, because each layer genuinely sits on the one below it.
The LLM is raw predictive capability — powerful, ungrounded, can’t act. An agent adds the ability to act, but stays generic, built for everyone. An AI Co-Worker adds the firm — your data, your reasoning, your way of working — and turns generic capability into something that’s actually yours.
Most of the confusion in the market is people comparing across layers as if they were the same purchase. Someone demos an agent and a firm thinks it just saw a Co-Worker. Someone hears “it runs on one of the big-lab models” and assumes the intelligence lives in the model, when the question that matters is what’s built on top of it. The whole point of this first post is to make those mix-ups visible, so that when a vendor shows you something, you can place it: which layer is this, actually?
That single question — which layer am I looking at — will save you more money than any other in your AI evaluation.
Where this series goes next
Now that the vocabulary is in place, the rest of the series builds up the stack in the order a firm actually has to build it.
Next: the data foundation — the medallion architecture, and why getting your data right has to come before any of the AI, no matter how much you want to skip ahead. After that: how a firm’s own reasoning gets codified on top of that foundation; how the whole operating layer fits together; and finally, the part regulated firms care about most — how you trust, audit, and stand behind any of it.
We’re starting with definitions because the firms that win with AI aren’t the ones who move first. They’re the ones who understood what they were buying. Let’s start there.