The deployment of automated software systems called AI agents has recently exploded. A November 2025 report by MIT Sloan School of Management and Boston Consulting Group found that 35 percent of surveyed businesses had already deployed AI agents, while one other 44 percent planned to implement agentic AI soon.
To know the basics and potential impacts 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, in addition to the underlying models and mechanisms that power agentic AI systems.
Q: What’s agentic AI and the way is it different from generative AI models like ChatGPT and Claude?
A: Agentic AI is AI that takes actions on the earth. These actions may very well be a physical motion, like robotic manipulation, or a digital motion, like booking a flight. Then again, we predict of generative AI as making up stories, poems, art, and pictures, relatively than taking actions for us.
The word “agent” is only a brand name. It normally means AI that’s going to assist people interact with an application, a web site, or the physical world. Most agents we encounter today are digital agents, like customer support agents you possibly can talk with about product complaints.
Most corporations that supply agents use the identical few AI models under the hood and provides them the flexibility to take actions and remember what happened. An agent starts with a fundamental generative AI system, like Claude, on the core. Then corporations put different wrappers around that foundation model for his or her product or application. Those wrappers could be specific tools that agent can use, and people tools depend upon the applying. Perhaps the agent has access to a calculator so it could solve math problems, or perhaps it has access to a more complicated hard disk drive and operating system so it could remember a firm’s financial data and past business negotiations.
The most important challenge in developing agentic AI comes from a scarcity of coaching data. If I would like to create a system that may log on and book a flight for me, that seems pretty easy. But we don’t have plenty of data that spells out exactly the best way to try this — where to maneuver the mouse, which buttons to click on, what to do if something goes unsuitable, or the best way to call any individual and negotiate concerning the price of the airline ticket. One approach to train a system like that is to have the AI agent visit airline web sites, try things out, and see what works and what doesn’t work. These environments are hard to model, so often the agent must learn by trial and error.
Q: What are some promising applications of agentic AI?
A: I believe the world where we’ve seen essentially the most success has been with coding agents. That is something that evolved from generative AI. People trained language models on code, after which they will predict what a human would do to unravel a coding problem. As well as, an agent can learn to do that by going through a feedback loop where it tries out different solutions and checks to see if it got the reply right. So long as it could check the reply, the AI agent can perform this trial-and-error loop until it figures out strategy.
But there may be all the time a balance between automating decision making versus simply assisting and informing humans. Analytical AI methods, just like the systems that help predict possible outcomes of choices, should not agentic in nature, but are very informative to human decision-makers. For cases which can be either high-stakes or safety-critical, like medicine, security, high-level business policies, etc., the technology won’t be ready for AI to completely automate those processes, or we won’t even be comfortable with that.
Q: Are there risks we needs to be fascinated about when using AI agents?
A: One big risk area comes from the undeniable fact that it is commonly very easy to get agents to do certain kinds of give you the results you want. With coding agents, you possibly can “vibe code” and just ask the agent to make a code for you, so that you don’t need to do the labor yourself. There may be a giant risk that, since it is really easy, people won’t put enough effort into verifying that it’s doing the best thing. Bugs can be introduced, private data will get leaked — that is already happening.
Agents aren’t perfect, within the sense that they may make mistakes because they should not well-trained and don’t know what to do. But even in the event that they are very competent, if a human doesn’t use them appropriately or gives them an instruction that is just too vague, the AI agent could make a mistake since the human made a mistake. If humans are less involved in considering through all the implications, I believe we could be more susceptible to making those mistakes.
A further aspect is the chance of de-skilling. It’s unclear how far it will go, but after we are counting on agents to do our homework, our coding, and our math, we would lose the flexibility to try this ourselves, and we would lose that ability too soon since the technology is just not yet ready to totally automate those processes.
Q: What does the long run hold for agentic AI?
A: What we predict of now as agentic AI refers to large language models using tools to interact with digital and physical systems. One obvious limitation is that, under the hood, these have the architecture of a language model and are trained on text data. To make much more powerful AI agents, we would must model videos, physical forces, time series, radar scans, and other modalities. We would must have models with fundamentally different architectures that may handle continuous data, high-dimensional data, stochastic data, and so forth.
But, however, perhaps an especially good coding model could act as a puppeteer to interface with sensors, actuators, and web APIs? Perhaps, once you will have a super-smart reasoning system that understands math, language, and code, you possibly can give it a camera and a keyboard and it’ll work out what to do within the spatial domain. Is the subsequent wave of AI just going to be Claude with sensors, actuators, and tools, or is it going to be something inbuilt a brand new way from the bottom up? That’s the large query plenty of people in AI are grappling with right away.