The use of automated software systems, so-called AI agents, has exploded recently. A November 2025 report from MIT Sloan School of Management and Boston Consulting Group found that 35 percent of companies surveyed had already deployed AI agents, while another 44 percent planned to implement agent AI soon.
Understand the fundamentals and potential impact of these increasingly 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 of AI agents and the underlying models and mechanisms that power agent AI systems.
Q: What is agentic AI and how does it differ from generative AI models like ChatGPT and Claude?
A: Agentic AI is AI that takes action in the world. These actions can be a physical action like robot manipulation or a digital action like booking a flight. On the other hand, we think that generative AI makes up stories, poems, art and images instead of taking actions for us.
The word “agent” is just a brand name. Typically, this involves AI that helps people interact with an application, a website, or the physical world. Most agents we encounter today are digital agents, such as customer service representatives who you can talk to about product complaints.
Most companies that offer agents use the same few AI models under the hood, giving them the ability to take action and remember what happened. An agent starts with a basic generative AI system like Claude at its core. Companies then place different wrappers around this basic model for their product or application. These wrappers can be specific tools that the agent can use, and these tools depend on the application. The agent may have access to a calculator to solve math problems, or perhaps he has access to a more complicated hard drive and operating system so he can memorize a company’s financial data and past business negotiations.
The biggest challenge in developing agent AI is the lack of training data. If I want to create a system that can go online and book a flight for me, it seems pretty easy. But we don’t have a lot of data that describes exactly how to do this – 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. One way to train such a system is to have the AI agent visit airline websites, try things out, and see what works and what doesn’t. These environments are difficult to model, so the agent often has to learn through trial and error.
Q: What are some promising applications of agent AI?
A: I think the area where we had the most success was coding agents. This is something that has evolved from generative AI. Humans train language models on code and can then predict what a human would do to solve a coding problem. Furthermore, an agent can learn this by going through a feedback loop where it tries different solutions and checks whether it gave the right answer. As long as it can verify the answer, the AI agent can go through this trial-and-error loop until it finds a good strategy.
However, there is always a balance between automating decision-making and simply supporting and informing people. Analytical AI methods, like the systems that help predict possible outcomes of decisions, are not agentic in nature but rather highly informative to human decision makers. In cases where the stakes are high or safety-critical, such as medical, security, high-level business policies, etc., the technology may not be ready for AI to fully automate these processes, or we may not even be comfortable with it.
Q: Are there risks we should think about when using AI agents?
A: A big risk is that it is often very easy to get agents to do certain work for you. Coding agents allow you to “vibe code” and simply ask the agent to create a code for you, so you don’t have to do the hard work yourself. Because it’s so simple, there’s a big risk that people won’t put enough effort into checking that it’s doing the right thing. Bugs will occur, private data will be leaked – it’s already happening.
Agents are not perfect in the sense that they may make mistakes because they are not well trained and do not know what to do. But even if they are very competent, if a human doesn’t use them properly or gives them too vague an instruction, the AI agent could make a mistake because the human made a mistake. If people are less involved in thinking through all the consequences, I think we may be more prone to making these mistakes.
Another aspect is the risk of deskilling. It’s unclear how far this will go, but if we rely on agents to do our homework, our coding, and our math, we may lose the ability to do it ourselves, and we may lose that ability too soon because the technology isn’t ready to fully automate these processes.
Q: What does the future of agent AI look like?
A: What we now call “agentic AI” refers to large language models that use tools to interact with digital and physical systems. An obvious limitation is that under the hood these have the architecture of a language model and are trained on text data. To develop even more powerful AI agents, we may need to model videos, physical forces, time series, radar scans, and other modalities. We may need models with fundamentally different architectures that can handle continuous data, high-dimensional data, stochastic data, etc.
But on the other hand, perhaps an extremely good coding model could act as a puppeteer for interfacing with sensors, actuators and web APIs? Once you have a superintelligent thinking system that understands math, language and code, maybe you can give it a camera and a keyboard and it will figure out what to do in the spatial domain. Will the next wave of AI be just Claude with sensors, actuators and tools, or will it be something built from the ground up? This is the big question that many people in AI are grappling with right now.
https://news.mit.edu/2026/agentic-ai-and-what-do-we-want-it-be-0630
