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

What the MIT Q&A Actually Reveals

Agentic AI has become one of the tech industry's favorite buzzwords, slapped onto everything from browser extensions to enterprise software suites. But a recent MIT News interview with Phillip Isola, an associate professor in EECS and a CSAIL researcher, attempts to cut through the marketing haze and ask a more basic question: what actually separates an "agent" from the chatbots we already know, like ChatGPT and Claude?

The distinction matters more than it might seem. Generative AI, in Isola's framing, is fundamentally about producing outputs — text, images, code — in response to a prompt. Agentic AI, by contrast, implies something closer to autonomy: a system that can take actions, make decisions across multiple steps, and pursue a goal with less direct human steering at each turn. That's a meaningful architectural and conceptual leap, not just a rebrand.

Why This Distinction Matters for Developers

For developer tools, the difference between generative and agentic systems is not academic. Building reliable agents requires new infrastructure: ways to let models call external tools, retain memory across steps, verify their own outputs, and recover from errors without human intervention at every juncture. This is why so much recent developer-tool investment — from coding assistants that can execute multi-step tasks to frameworks for orchestrating chains of AI actions — is racing to catch up with a vision of "agentic" behavior that, as Isola's comments suggest, is still more aspirational than fully realized.

That gap between the marketing term and the technical reality is worth sitting with. If agentic AI today is largely generative AI wrapped in scaffolding that lets it take sequential actions, then developers evaluating these tools should be cautious about assuming genuine independent reasoning or judgment. The underlying models are still prone to the same hallucinations, brittleness, and lack of true understanding that plague their generative predecessors — agentic wrapping doesn't automatically fix that.

Consumer Expectations Are Running Ahead of the Technology

For everyday users, the stakes are arguably higher. Consumer products are increasingly pitched as having "agents" that can book travel, manage schedules, or shop autonomously. If the underlying reality is closer to sophisticated automation than genuine independent intelligence, consumers risk over-trusting systems that can fail silently or make consequential errors — booking the wrong flight, spending money incorrectly, or leaking sensitive data during multi-step tasks.

The Bigger Question: What Do We Want This to Be?

Perhaps the most useful part of framing this as "what is agentic AI today, and what do we want it to be" is the acknowledgment that the field is still being defined. That leaves room — and responsibility — for researchers, companies, and regulators to shape agentic AI deliberately, rather than let the term's popularity outpace its substance.