Close-up Portrait of Software Engineer Working on Computer, Line of Code Reflecting in Glasses. ... [+] Developer Working on Innovative e-Commerce Application using Machine Learning, AI Algorithm, Big DatagettyIf youve been paying attention to industry chatter, youll know that people are talking a lot about the era of genetic AI. Were not used to that word, agentic, so maybe I should say that were discussing the role of AI agentsTo those unfamiliar with recent research, though, its somewhat unclear. What is an AI agent, and how does it differ from what came before with LLMs and neural networks?Ive come up with a rather simple way to explain how AI agents work, based on what people in the MIT media lab and elsewhere are telling us.Three Pillars of AI Agent BehaviorsFirst, AI agents can perceive things about the world around them.This idea of AI perception starts with computer vision, but it doesnt end there, either. If an AI agent is doing a task, like, say, booking a flight on behalf of a human, it has to perceive a large number of things about its digital environment where the booking engine is, how to use it, what each set of symbols means, etc.The second big idea is that AI agents have to be able to reason, to some extent. Simple LLM chatbots are just generating outputs. But agents need to make decisions they need to be able to figure out whats in the best interest of the humans they are working for.The third point is perhaps the most interesting about AI agents they need to know how to act.Now, what does this mean? It means that these AI entities are not just waiting for a user-generated event. Theyre not waiting for someone to ask a question, or give them a command. Theyre acting on their own, in real time, based on a goal or instruction.I think that key phrase is important: in real time. If you ask an AI to do something and it does it in the moment, thats more of a linear deterministic response.If, on the other hand, the AI is waiting in the background to do its work, figuring out the best time for any given action, that adds another dimension to its capabilities.Young Technologists at DavosA recent panel at Davos really explained this in a good way. These are young researchers from the MIT media lab and the private sector who understand the work that people are doing now and how AI agents are different.Responding to questions from the moderator, Tobin South, James Rubin gave this definition of an AI agent, contrasting it with chatbots, which, he said, just generate a single Turing prediction.An agent can act on your behalf, he said. So whereas a chatbot may just plan out a list of tips on how to book travel, an agent could actually go about using external tools to do those things. And the way it really goes about doing this is, it's given a goal, for example, book a trip. And (then it) engages in this kind of reasoning, and actually (uses a) reflection loop iteratively, you know, continuing that until it reaches an acceptable outcome.Hope Schroeder talked about how AI can help us to either counteract or reinforce our own biases, and Kevin Dunnell described AI agents as prophecy machines that can give us prompts toward data-driven insights.Looking to the FutureEach of these young data scientists had some words in responding to South asking them what they would tell people who are reacting to the latest AI news.Schroeder encouraged the audience to get curious about what AI can do.You're empowered by the information you're able to get about language models and AI, the more you're able to engage with it and actually think about what the opportunities and limitations are, she said. It's very easy to feel threatened and disempowered by some of the changes that are happening as a result of AI to feel disempowered in work, let's say, through automation, or a sense that AI might be taking your job, or disempowering you creatively, or in your decision making. But I would encourage people who feel that sense of disempowerment to get curious about what the system is, what its limitations are, and as you empower yourself with that information, you might be able to be part of shaping the conversation around how to make this better, and help turn the tide towards allowing it to have a empowering impact, rather than a disempowering impact.Rubin suggested that people should take both the doom and the euphoria with a grain of salt, noting that with all of what AI can do, it still cant write syntactically correct APIs.People should continue to have these types of discussions, so that we can shape it incrementally in the right direction, he said.Using the example of a Cartesian system, Dunnell urged people to take advantage of the capabilities that AI now offers.We're just about to be able to start seeing qualitative information, which is much more complex than this quantitative information that we have, he said. And I see AI as a tool that makes that possible for us, and that ultimately, the value of that is what we can do with that information as humans, as opposed to what that system can do.All of this highlights the realities that were dealing with in moving into the future with AI systems. Were going to have to understand how these machines work as we try to integrate them into our lives and our businesses. Knowing their capabilities, that these engines can act in real time on their own reasoning, is crucial.