For the last year we organized these issues around themes and stories. Though incredibly fun, planning the next theme will take a bit of time. Thus, this season we are doing something simpler. We’re picking ideas common in AI that people ask us about regularly that are worth understanding. We kick off this season with the word anyone who is remotely connected to AI can’t get away from right now — agent.
Welcome to the New Season
If you have spent any time around AI in the last year, you have heard that "agents" are coming, that this is "the year of the agent”, that agents will book your travel, run your clinic’s back office, and write your notes while you sleep. The word “agent” is doing a lot of work in common parlance, and most of the time nobody stops to say what it means.
That is a problem because “agent” is not marketing jargon invented last Tuesday. It is one of the oldest ideas in AI and it maps well onto how behavior analysts already think. Once you see the structure underneath an agent, most of the hype becomes easy to parse, and the stuff worth taking seriously becomes easier to spot.
This week we review what an agent actually is, where the idea came from, what people mean by it today, and how it connects to subagents (a newer term you are starting to hear carrying the spectre of ouroboros).
The Core Idea is A Loop, Not an Answer
The entire concept behind an agent is simple. A chatbot gives you an answer. An agent takes actions in a loop until a goal is met.
When you ask a chatbot a question, it responds once and stops. One input, one output. An agent works differently. The system receives context, chooses a next action, takes that action, observes what happened, and then decides what to do next. And, this loop simply repeats over and over until the task is finished or it hits a stopping rule.
That loop is the entire difference. People in the field only half-joke that an agent is “a language model in a while loop with tools”. One popular shorthand puts it as: Agent = model + memory + planning + tool use. The model is the part that decides (i.e., map an input to an output); the tools are the hands (doing the actual work because LLMs are not very skilled); the loop is what lets it keep going instead of stopping after one reply.
Behavior analysts should feel a flicker of recognition here, because this is a feedback loop, and feedback loops are our native language. An organism does something, contacts a consequence, and its next action is a function of what just happened. Antecedent, behavior, consequence—and then around again. The agent loop has the same shape. Situation, action, result, next action. The physical machinery is different and the comparison is structural, not literal (there is no organism, no motivating operation, no reinforcement in the technical sense). But the reason agents feel intuitive to us is that we have been reasoning about closed behavioral loops our entire careers.
A Short History: The Word Is Older Than the Hype
The idea of an “agent” long predates ChatGPT. It is worth knowing the lineage, because it tells you what has actually changed recently as well as what hasn't.
Classical AI. In the standard textbook framing that shaped decades of AI, an intelligent agent was anything that could perceive its environment through sensors and act on it through actuators. That is deliberately broad. A thermostat is a (very simple) agent, and so is a self-driving car. Researchers sorted them by how much complexity was going on in the middle. Did the system work fine with simple reflexes that just map a simple input to an output? Or was that insufficient and the system needed to independently derive goals and learn how to adjust to the world over time? Regardless, the organizing question was always the same. Given what it senses, what should it do to move toward its goal?
