In Game Theory and RL we talk a lot about agents. We might imagine that an agent is a magical shining sphere floating in space, and inside there's something like a real person. Or an animal. Or an AI algorithm. Anything that cares about doing actions that will get it as much reward as possible. The agents are the "people" roaming through the wild terrain of states, observations, actions and rewards, looking for their happy place.
Mathematicians do like mathematicians be. In any field, they like to set up their axioms and then build towers of knowledge on top of them. They draw a line in the sand, separating the questions they want to answer from the questions they'd like to leave to other people to each find their own. The location and shape of that border defines the field and its character. Understanding the border is essential to understanding the field.
What does that border look like in Game Theory?
Game Theorists decided that they *would* like to answer questions like:
1. "What decision tree of actions that the agent can take will result in the highest reward?"
2. "Under which behavior of agent 1 will agent 2 be able to get the highest reward?"
3. "Is there a combination of actions for different agents, such that: Assuming all the agents but a specific agent choose actions from the combination, the specific agent will get the highest reward for also taking the action of the combination?"
Game Theorists decided that they *would not* like to answer questions like:
1. "What is an agent?"
2. "Does an agent *want* to get a high reward? What does 'want' mean?"
3. "What decision tree of actions that the agent will take will make it the happiest?"
## What is an agent, really?
When you ask "what is an agent?" you're almost asking: "What is a living creature? What is a soul?" and as long as we're there why not tack on "what is the meaning of life?" No wonder Game Theorists steer far and away from this territory. I won't make a dent in the "meaning of life" question in this article. But maybe I'll get some clarity on the question: "What do we consider to be alive in Game Theory?"
A naive view would be that an agent is anything smart enough to make decisions, like a person or a rat or an AI agent. When a swirling soup of atoms achieves Turing-completeness, an agent pops up in the center of it.
I don't think that interpretation will get you far. Life isn't a video game.
When we call a collection of matter an "agent" it's more about our attempt to understand the way it behaves. Life is the phonemenon we're studying and "agent" is a model that we use for it. The phenomenon is not the model. The model is a toy that we use to explain the phenomenon and make predictions about its behavior.
When we say that a fruit fly is an agent, we say "we decide to now use the model called 'agent' on this fly, so we could explain its past actions and make predictions of its future actions." We make this decision for every piece of matter that we encounter. **When we say that something is an agent, we can imagine ourselves drawing a line around that piece of matter, separating the agent from its world.** We can call that our "agent definition". We have many agent definitions for different pieces of matter in the world. The sum of these agent definitions is like a map with multiple closed shapes on it, that are drawn around different chunks of matter. This "agent map" defines our perception of the world.
This philosophical game is made more enchanting by the following observations:
1. **Different people will have different agent maps.** You might think that a single-celled organism is an agent while I might think it's a simple automaton. I might think that The Coca Cola Company is an agent while you would think that it's a collection of smaller agents, which are its employees. People can also change their minds on their agent definitions, or have several different viewpoints, i.e. several different agent maps that they pull out in different situations.
2. **Agent definitions of different people might be overlapping**. The difference between agent maps can't be expressed as a list of all the agents that the first map defines and the second doesn't, because the second agent map might have an agent definition that completely or partially contains matter from an agent definition of the first agent map. An example of partial containment is when I have an agent definition called "The Coca Cola Company" while you have an agent definition "The American People." There are pieces of matter that are members of both sets, and those are the American employees of The Coca Cola Company. My agent shares a few neurons with your agent.
3. **The people who decide what is an agent, are actually agents themselves.** Here we come full circle. We use the term "agent" to describe an entity that's similar to us. In the same way that we draw these circles around matter in the universe, some of that matter in the universe is drawing circles around us.
4. **There are no "correct" or "incorrect" agent definitions.** An agent map is simliar to a frame of reference in the world of Newtonian Physics. We can describe one frame of reference as stationary and another as moving, but we'll never have one true answer to which of the frames of reference is actually at rest. Similarly, there isn't an objective answer to the question "Is chunk of matter X an agent or not?" There is only "Who are the other chunks of matter that choose to view X as an agent, and why do they choose to do that?"
That last thought is the most pertinent. Whenever we find ourselves asking "Is this an agent?" we can draw down the shades and ask, _"Do we want it to be?"_
Let's dive into this second question.
## Why do we decide to view something as an agent?
When we hear the phrase "Microsoft is hiring", we think of Microsoft as an agent. Microsoft is a company that would like to expand, i.e. encompass more human matter, and it's now taking steps to lure those hires in. If we're on a hiring committee in Microsoft making a decision on a candidate, we are likely to have a more fine-grained view. We wouldn't be so interested in the fact that Microsoft, the big agent, wants to expand. We'll be thinking: "Do I want to work with this guy? Will this candidate take our team in a direction that we want to go? Do I want to express my opinion about this candidate to my peers on the committee with confidence, or do I prefer to defer to other people's opinions and let them set the frame of discussion?"
In other words, we'll be viewing each of the different people on the comittee, and the candidate, as separate agents. When we're on the hiring committee, we need to understand how the people in it behave and how we can succeed. For that, this fine-grained view provides better explanations and makes better predictions about what will happen with these people than the coarse-grained view that sees Microsoft as one big, single agent.
The single-agent view isn't more or less correct than the multi-agent view. They're just different tools made for different jobs.
todo: people's political differences can be viewed as differences in their agent maps. maybe it should be a separate article.