Phillip Isola, associate professor in MIT’s Department of Electrical Engineering and Computer Science and a member of CSAIL, has published a clear-eyed breakdown of what agentic AI actually is, where its technical limits sit, and what risks come with rapid adoption. The Q&A, published by MIT News, cuts through marketing language to define the category in engineering terms.
“Agentic AI is AI that takes actions in the world,” Isola told MIT News. “The word ‘agent’ is just a brand name. It usually means AI that is going to help people interact with an application, a website, or the physical world.”
The Architecture Under the Label
Isola’s technical breakdown is direct: most commercial agents wrap one of a few foundation models (Claude, GPT, Gemini) with action-taking capabilities and memory. “An agent starts with a fundamental generative AI system, like Claude, at the core. Then companies put different wrappers around that foundation model for their product or application,” he said.
Those wrappers include domain-specific tools (a calculator, a file system, an API) and memory systems that let the agent retain context across interactions. The distinction from a stateless chatbot is persistent state and the ability to execute actions, not a fundamentally different model architecture.
The Training Data Bottleneck
The most technically substantive observation in Isola’s analysis is about what holds agents back. “The biggest challenge in developing agentic AI comes from a lack of training data,” he said. Booking a flight seems simple, but the training data for that task, including where to move a cursor, which buttons to click, what to do when errors occur, is sparse compared to the text corpora used to train language models.
The workaround is trial-and-error reinforcement: agents visit websites, try actions, and learn from outcomes. This works well in domains where answers are verifiable (coding agents can test their output), but poorly in domains where correctness is ambiguous or consequences are irreversible.
Two Risks Worth Tracking
Isola identified two risks that map to active deployment patterns. First, verification failure: “People will not put enough effort into verifying that it is doing the right thing. Bugs will be introduced, private data will get leaked. This is already happening.”
Second, de-skilling: “When we are relying on agents to do our homework, our coding, and our math, we might lose the ability to do that ourselves, and we might lose that ability too soon because the technology is not yet ready to fully automate those processes.”
The adoption numbers suggest these risks are already materializing at scale. A November 2025 MIT Sloan/BCG report found 35% of surveyed businesses had deployed agents, with 44% planning implementation soon.
The Open Architectural Question
Isola closed with a question that matters for anyone building on top of current agent infrastructure: “Is the next wave of AI just going to be Claude with sensors, actuators, and tools, or is it going to be something built in a new way from the ground up?”
If the answer is the former, current agent frameworks and orchestration layers are building on stable ground. If the answer is the latter, significant architectural rewrites are ahead. Isola does not predict which path wins, but he frames the stakes clearly: the foundation model wrapper approach has scaling limits that may require fundamentally different model architectures for physical-world and continuous-data domains.