MIT professor Phillip Isola explains what agentic AI actually is

Computer scientist Phillip Isola, an associate professor in MIT's Department of Electrical Engineering and Computer Science (MIT CSAIL), sat down with MIT News to explain what the term agentic AI actually means and how the real capabilities of these systems differ from the marketing hype surrounding them.

What is agentic AI

A regular chatbot simply responds to a user's message. An AI agent, by contrast, can plan a sequence of actions on its own: it breaks a task into steps, calls external tools, browses web pages, or runs code, and it adjusts its plan as it goes. This ability to act autonomously across multiple steps, rather than just generate text, is what sets agentic AI apart from classic large language models.

• Interviewee: Phillip Isola, associate professor at MIT CSAIL
• Topic: what agentic AI is today and what researchers want it to become
• The piece ran in the MIT News section dedicated to AI research

How agents differ from the hype

Over the past two years, the industry has shipped dozens of products branded as "AI agents" — from browser assistants like OpenAI's Operator to Anthropic's computer-use features and autonomous coding agents like Devin. But they are far from equally autonomous. Some systems only imitate agency through rigidly scripted workflows. Researchers like Isola are focused on a more fundamental question: how to teach a model to genuinely plan and adapt to new situations instead of just executing a predefined sequence of actions.

Isola himself specializes in computer vision and generative models, and is known, among other things, for his work on image-to-image translation algorithms that convert images between styles. That background in fundamental research on generative systems lets him view the "agent" craze from the outside — not as a product marketer, but as a researcher for whom the line between "the model responds" and "the model acts autonomously" is a matter of rigorous technical analysis, not an advertising slogan.

Where the field is heading

The key question posed by the MIT News headline is not only "what is agentic AI today" but also "what do we want it to be." That framing makes clear that building agentic systems is a normative challenge as well as an engineering one: the research community still has to determine what level of autonomy is safe and useful before these systems get more independence in real-world tasks — from booking tickets to managing finances or production code.

What it means

As the market fills up with products carrying the "agent" label, the sober view from MIT researchers is a useful reminder: true AI autonomy is an open research problem, not a solved technology you can simply wrap in an interface and sell.