The deployment of automated software programs called AI agents has recently exploded. November 2025 report by MIT Sloan School of Management and Boston Consulting Group found that 35 percent of surveyed businesses have already deployed AI agents, and another 44 percent plan to implement agent AI in the near future.
To understand the basics and potential implications of these popular toolsMIT News spoke with Phillip Isola, an associate professor in 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 basic models and methods that power AI systems.
Question: What is agent AI and how does it differ from generative AI models like ChatGPT and Claude?
A: Agentic AI is AI that performs actions in the world. These actions can be a physical action, such as manipulating a robot, or a digital action, such as booking a flight. On the other hand, we think of productive AI as creating stories, poems, art, and images, rather than taking action for us.
The word “agent” is just a brand name. It usually means AI that will help people interact with an app, website, or virtual world. Most of the agents we meet today are digital agents, like customer service agents you can talk to about product complaints.
Many companies that provide agents use several similar AI models under the hood and empower them to perform actions and remember what happened. The agent starts with a basic AI generator, like Claude, at the end. Then companies put different wrappers on that base model of their product or application. Those wrappers may be specific tools that the agent can use, and those tools are application dependent. Perhaps the agent has access to a calculator to solve mathematical problems, or perhaps he has access to a sophisticated computer and operating system to recall the firm’s financial data and past business transactions.
A major challenge in developing agent AI comes from the lack of training data. If I want to create a system that can go online and book a flight for me, that seems pretty easy. But we don’t have a lot of data that explains exactly how to do that — where to move the mouse, what buttons to click, what to do if something goes wrong, or how to call someone and negotiate the price of a plane ticket. Another way to train a system like this is to have an AI agent visit airline websites, try things out, and see what works and what doesn’t. These areas are difficult to match, so the agent must learn by trial and error.
Question: What are some promising applications for agent AI?
A: I think the area where we’ve seen the most success has been with coding agents. This is something that came from AI for manufacturing. People who are trained to model language in code, then cannot predict what a person might do to solve a coding problem. Furthermore, the agent can learn to do this by using a feedback loop where it tries different solutions and checks to see if it got the answer right. As long as it can evaluate the answer, the AI agent can do this trial and error loop until it finds a good strategy.
But there is always a balance between making automatic decisions versus just helping and informing people. AI analytical methods, such as programs that help predict the outcomes of potential decisions, are not agency in nature, but rather instructive to human decision makers. In potentially high-risk or safety-critical situations, such as medicine, safety, high-level business policies, etc., the technology may not be ready for AI to automate those processes, or we may not be comfortable with that.
Question: Are there any risks we should think about when using AI agents?
A: One major area of risk comes from the fact that it is often very easy to find agents to do certain types of work for you. With code agents, you can “vibe code” and simply ask the agent to code for you, so you don’t have to do the hard work yourself. There’s a big risk that, because it’s easy, people won’t put enough effort into making sure it’s doing the right thing. Bugs will be introduced, private data will be leaked – this is already happening.
Agents are not perfect, in the sense that they can make mistakes because they are not properly trained and do not know what to do. But even if it is very skilled, if a person does not use it correctly or gives it a command that is not very clear, the AI agent can make a mistake because a person made a mistake. If people are less involved in thinking through all the consequences, I think we might be more prone to making those mistakes.
An additional factor is the risk of de-skilling. It’s unclear how far this will go, but if we rely on agents to do our homework, coding, and math, we may lose the ability to do that ourselves, and we may lose that ability very soon because the technology isn’t ready to automate those processes.
Question: What is the future of agent AI?
A: What we now think of as agent AI refers to large-scale language models that use tools to communicate with digital and physical systems. Another obvious limitation is that, under the hood, these have language model properties and are trained on text data. To make AI agents even more powerful, we may need to model videos, physical forces, time series, radar scans, and other methods. We may need to have models with very different properties that can handle continuous data, high-dimensional data, stochastic data, and so on.
But, on the other hand, maybe the best coding model could act as a puppeteer to communicate with sensors, actuators, and web APIs? Maybe, if you already have a very intelligent thinking system that understands math, language, and code, you can give it a camera and a keyboard and it will figure out what to do in the local domain. Will the next wave of AI be just Claude with sensors, actuators, and tools, or will it be something completely new from the ground up? That’s the big question many people in AI are facing right now.
