Q&A: What is agentic AI today, and what do we want it to be?

What the MIT Q&A Actually Addresses

A recent MIT News Q&A featuring Phillip Isola, an associate professor in EECS and a CSAIL researcher, tackles a question that has become surprisingly hard to pin down as AI marketing accelerates: what actually distinguishes 'agentic AI' from the generative chatbots most people already use. Isola, whose research focuses on the intelligence underlying AI agents and the mechanisms that drive agentic behavior, uses the conversation to draw a clearer line between systems like ChatGPT or Claude that respond to prompts, and agentic systems designed to take multi-step actions, make decisions, and pursue goals with less direct human steering.

Why the Distinction Matters

The generative-versus-agentic distinction isn't just semantic housekeeping — it has real consequences for how these systems are evaluated and governed. Generative models are largely judged on the quality of a single output: is the text coherent, is the code correct, is the image accurate. Agentic systems, by contrast, chain together reasoning, tool use, and autonomous decision-making over extended tasks, which means errors can compound, goals can be misinterpreted, and behavior can diverge from what a human operator intended across many steps rather than one.

This has direct implications for AI safety research and alignment work. Evaluating a single response for harmful content is a fundamentally different problem than evaluating whether an agent will pursue a flawed sub-goal, take an unsafe action in the world, or manipulate its environment to complete a task more efficiently. As agentic systems increasingly get access to tools, code execution, web browsing, and other real-world levers, the surface area for things to go wrong expands considerably.

Implications for Evaluation and Red Teaming

Frontier model evaluation frameworks built for chatbots — benchmarks around factuality, toxicity, or refusal behavior — don't automatically transfer to agentic contexts. Red teamers probing agentic systems need to test not just outputs but action sequences: does the agent take unauthorized steps, escalate privileges, or pursue a goal in ways its designers never anticipated? This is precisely the kind of question academic researchers like Isola are positioned to help formalize, since much of current red-teaming practice for agents is still ad hoc, borrowed from generative-model testing, or specific to individual companies' internal processes.

The Bigger Picture

As the industry pushes agentic AI into coding assistants, customer service, and autonomous research tools, having academically grounded definitions matters for regulators, safety researchers, and the public trying to assess risk. Clarifying what agentic AI is — and isn't — is a prerequisite for building meaningful safety evaluations, rather than retrofitting old benchmarks onto fundamentally different systems.