Computer scientist Phillip Isola cuts through the hype to describe how AI agents work and what the future might hold for this fast advancing technology.
The deployment of automated software systems known as AI agents has currently exploded. A November 2025 report by MIT Sloan School of Management and Boston Consulting Group found that 35% of surveyed businesses had already deployed AI agents, even as some another 44% intended to execute agentic AI quickly.
To understand the fundamentals and potential effects of those increasingly popular tools, MIT News spoke with Phillip Isola, an associate professor within the Department of Electrical Engineering and Computer Science (EECS) and a member of the Computer Science and Artificial Intelligence Laboratory (CSAIL), who studies the intelligence AI agents possess, as well as the underlying models and mechanisms that power agentic AI systems.
Q: What is agentic AI and how is it exclusive from generative AI models like ChatGPT and Claude?
A: Agentic AI is AI that takes in the world. These actions may be a physical action, like robotic manipulation, or a digital action, like booking a flight. On the other hand, we think of generative AI as composing up stories, poems, art, and images, instead of taking action for us.
The word “agent” is just a brand name. It generally means AI this is going to assist people engage with an application, a website, or the physical world. Most agents we come across today are digital agents, like customer service agents you can talk with about product complaints.
Most companies that provide agents use the same few AI models under the hood and give them the ability to take actions and consider what happened. An agent begins with a fundamental generative AI system, like Claude, at the core. Then companies positioned different wrappers around that foundation model for their product or application. Those wrappers is probably unique tools that agent can use, and those tools rely upon the application. Maybe the agent has access to to a calculator so it can solve math problems, or perhaps it has access to a more difficult hard drive and working system so it can remember a firm’s financial data and past business negotiations.
The biggest task in developing agentic AI comes from a lack of training data. If I need to create a system which could go online and book a flight for me, that appears pretty simple. But we don’t have a lot of data that spells out precisely how to do that — where to move the mouse, which buttons to click on, what to do if something goes wrong, or how to call someone and negotiate approximately about the price of the airline ticket. One way to train a system like this is to have the AI agent visit airline websites, try things out, and notice what works and what doesn’t work. These environments are hard to model, so frequently the agent ought to learn by trial and errors.
Q: What are some promising applications of agentic AI?
A: I think the area wherein we’ve seen the most success has been with coding agents. This is something that developed from generative AI. People trained language models on code, after which they can expect what a human would do to resolve a coding issue. In addition, an agent can learn to do this by going through a feedback loop wherein it tries out other solutions and checks to see if it got the answer proper. As long as it can check the answer, the AI agent can carry out this trial-and-errors loop till it figures out a terrific method.
But there is always a balance among automating decision making versus simply supporting and informing humans. Analytical AI strategies, like the systems that help expect viable results of decision, aren’t agentic in nature, but are very informative to human decision-makers. For cases which are either high-stakes or safety-critical, like medicine, security, high-level business regulations, and so on., the technology might not be prepared for AI to completely automate those approaches, or we might not even be comfortable with that.
Q: Are there risks we need to be thinking about when using AI agents?
A: One big risk area comes from the facts that it’s miles easy very to get agents to do certain types of work for you. With coding agents, you can “vibe code” and simply ask the agent to make a code for you, so you don’t have to do the hard work yourself. There is a big risk that, due to the fact it’s so easy, people will not put enough effort into verifying that it’s far doing the right things. Bugs may be launched, private data will get leaked — this is already going on.
Agents aren’t perfect, in the sense that they might make mistakes because they are not properly-trained and don’t know what to do. But even if they’re very competent, if a human doesn’t use them accurately or gives them an instruction that is too vague, the AI agent could make a mistake due to the fact the human made a mistake. If humans are less concerned in wondering through all of the consequences, I think we might be more prone to making those mistakes.
An additional component is the risk of de-skilling. It is unclear far this will go, however when we are depending on agents to do our homework, our coding, and our math, we might lose the ability to do that ourselves, and we might lose that ability too soon because the technology is not ready to fully automate those techniques.
Q: What does the future hold for agentic AI?
A: What we think of now as agentic AI refers to large language models using tools to interact with digital and physical systems. One apparent limitation is that, under the hood, those have the architecture of a language model and are trained on text data. To make even more powerful AI agents, we might need to model videos, physical forces, time serious, radar scans, and other modalities. We would possibly want to have models with basically different architectures that can cope with persist records, high-dimensional data, stochastic records, and so on.
But, on the other hand, maybe an first-rate coding model ought to act as a puppeteer to interface with sensors, actuators, and web APIs? Perhaps, as soon as you have a super-smart reasoning system that knows math, language, and code, you can give it a camera and a keyboard and it will figure out what to do in the spatial domain. Is the next wave of AI simply going to be Claude with sensors, actuators, and tools, or is it going to be something built in a new way from the ground up? That’s the huge question a lot of people in AI are grappling with right now.
