Q&A: Understanding Today’s Agentic AI and the Future We Want It to Become
Agentic AI acts via tools beyond text. Coding agents excel, but risks include errors, privacy leaks, de-skilling, and limits on human judgment.
Key Takeaways
- Agentic AI differs from generative AI because it takes actions, such as booking flights or manipulating software systems, rather than only producing text or images.
- Many current agents are built around foundation models like Claude or ChatGPT, with added tools and memory that determine what the agent can do.
- Coding agents are the clearest success because code allows fast feedback loops through testing, while tasks like real-world navigation or negotiation remain harder due to sparse training data.
- Major risks include mistakes from incomplete training or vague instructions, security and privacy issues from unchecked outputs, and long-term de-skilling if people rely too heavily on agents.
- The practical takeaway is to treat agents as powerful assistants: clarify what actions they can take, what feedback they can verify, and who must approve outcomes that affect money, data, health, safety, or business decisions.
Meta description:
MIT News asks Phillip Isola what agentic AI is, how it differs from generative AI, why Claude and ChatGPT matter, and what risks users should watch.
Tags: agentic AI, Claude, generative AI, AI agents, MIT News
Featured image: abstract, friendly illustration of a digital agent linking a chat bubble, tools, and a flight-booking interface; no real people, no logos, no text.
For readers deciding whether to let software systems act on their behalf, a November 2025 report from the MIT Sloan School of Management and Boston Consulting Group gives a clear signal: 35 percent of surveyed businesses had already deployed AI agents, while another 44 percent planned to implement agentic AI soon. MIT News spoke with Phillip Isola, an associate professor in the Department of Electrical Engineering and Computer Science and a member of the Computer Science and Artificial Intelligence Laboratory, to explain what these tools are, where they work, and where human judgment still matters.
The difference is action
Agentic AI is software that takes actions, not only produces text. It can move physically, such as robotic manipulation, or digitally, such as booking a flight. Generative AI creates stories, poems, art, or images for a person to read, view, or edit. ChatGPT and Claude are familiar generative systems; an agent may use one as its core reasoning component, then add tools and memory.
Isola says “agent” is largely a brand name. It usually means AI that helps people interact with an application, a website, or the physical world. Most agents encountered today are digital, like customer service agents for product complaints.
Many products use a small set of foundation models, such as Claude, and add wrappers. A wrapper may give the agent a calculator for math problems, or access to a hard drive and operating system so it can remember a firm's financial data and past business negotiations. The wrapper determines what the agent can do.
Coding agents are the clearest success
Coding agents evolved from generative AI models trained on code. They can predict what a human programmer would do to solve a problem. They also benefit from a feedback loop: try a solution, test it, check the result, and repeat until the strategy works. This loop is possible because code often has a clear pass-or-fail check.
Flight booking is harder. The outcome sounds simple, but the agent must know where to move the mouse, which buttons to click, what to do if something goes wrong, and how to call someone and negotiate a price. Training data for those steps is scarce. Isola describes agents visiting airline websites, trying actions, and learning by trial and error because real environments are hard to model.
Automation has limits
Analytical AI methods that predict outcomes are useful to human decision-makers, but they are not agentic because they do not act. Isola points to medicine, security, and high-level business policy as areas where full automation may not be ready, and where people may not want to give up judgment.
The risk is not only that agents make mistakes. They can fail because training is incomplete, or because a human gives a vague instruction. Coding agents make this visible: when “vibe coding” makes it easy to ask for code, users may skip checking the output. Bugs can enter products, and private data can leak. Isola says this is already happening. If humans think through consequences less, they may be more prone to errors.
A longer-term concern is de-skilling. If people rely on agents for homework, coding, and math before the technology can fully replace human judgment, they may lose the ability to do those tasks themselves.
What comes next
Today's agentic AI often means large language models using tools to interact with digital and physical systems. The limitation is architecture: these models are built on language-model designs and trained mostly on text. More powerful agents may need models that handle video, physical forces, time series, radar scans, and other continuous, high-dimensional, or stochastic (randomized) data, possibly requiring new architectures.
Another path is possible. A very strong coding model, paired with sensors, actuators, and web application programming interfaces, might act like a puppeteer for the physical world. If a reasoning system understands language, math, and code, it might learn to operate a camera and keyboard well enough to handle spatial tasks. Isola frames the open question: is the next wave simply Claude with added tools, or something built from the ground up?
For readers, the practical takeaway is to treat agents as assistants with real power. Ask what action the agent can take, what feedback it can check, and who must approve the result before it affects money, data, health, safety, or business decisions.
Comments (0)