Agency = Intelligence × Action. If either of those factors are zero then we don't have anything. When both of these factors exceed human level by a magnitude then we effectively have a Genie without a wish limit. So far Intelligence ≈ 1 if we use human level intelligence as a baseline while Action ≈ 0.1. So current agents are a poor substitute for humans but since their intelligence matches ours they can be as useful as a human in the limited domains where their action ability matches ours.

Since the Agency equation is identical to the productivity/money equation the company that makes human level action ability will have something that competes with humans for money. But by the time they get to human level action ability, intelligence will be superhuman. So these agents will easily outcompete humans making them redundant. Though outcompete is kind of the wrong word because humans reason for existing isn't to produce. It's to create beauty. And since we are the judges of beauty then by definition agents can't compete with us on that.

Many things we call beautiful today will cease to be beautiful. Much like ornate buildings ceased to be beautiful when we gained to ability to produce ornate finishing in factories. As such things that we consider beautiful today like movies and computer games will lose their beauty and become tools. Movies will become simply a communication medium and will be considered no more beautiful than an audiobook is today. The story underlying the movie will remain beautiful only if the characters inside the story are interesting enough for the consumer to want to interact with them. The stories of the future will be closer to simulations than stories since the initial pass through the simulation will be like a trailer for the simulation. You will watch a full length movie for free, with the producer hoping to get you interested in the characters and environment enough to purchase¹ the simulation.

So the end result is going to be beautiful, but it's obviously going to be extremely disruptive while these things are getting developed. In order to predict how this transition will play out we need to understand how an agent works. Because it's this that determines which industries get disrupted first.

An agent is a prediction function + memory + a physical body. The prediction function is what's currently known as an LLM. But that name will change when they warrant it. A model is just function from one world state to another. In order to create a significant difference between the input and output we need to combine the model with memory. This allows it to think through a problem step by step. Effectively allowing a less intelligent model to make leaps of insight as large as a smarter model. The memory also makes them easier to use since it removes the need for the user to provide all the required input in one go. They can build up input slowly. When you iterate on a project with Claude code you are essentially adding input until eventually the agent is able to predict what the exact thing is that you want and then it can finally give it to you. In this case most of the memory is the files that make up your project. Just because it doesn't all fit into the model at once make it any less input. Skills are just the input that is reused across projects.

The physical body part of the bit that's currently lacking the most. Though I don't think it's necessarily that hard. It's just that until now we weren't sure we would even be able to produce human level intelligence. And without that a great physical body doesn't make for a great agent. According to the equation we need both parts.