Artificial intelligence is entering a new phase with the emergence of agentic AI, systems designed not only to generate content but also to take actions in digital and physical environments. In an interview with MIT News, Phillip Isola, an associate professor in MIT’s Department of Electrical Engineering and Computer Science and a member of CSAIL, explains that AI agents differ from conventional generative AI because they can use tools, remember previous interactions, and complete multistep tasks on behalf of users.
According to Isola, most AI agents are built by extending existing large language models with specialized capabilities. These additions may include access to calculators, databases, operating systems, or web applications that allow the systems to perform practical tasks such as booking flights, assisting customers, or writing software. Rather than relying solely on text generation, agentic AI combines reasoning with action, making it useful for automating real-world workflows.
One of the most successful applications to date is software development. Coding agents can generate, test, and refine code through trial and error, using automated feedback to improve their performance. However, Isola cautions that the same approach is less suitable for high-stakes fields such as medicine, security, or business policy, where human oversight remains essential because mistakes can have serious consequences.
The interview also highlights several risks associated with increasing reliance on AI agents. Users may become less diligent about verifying AI-generated work, potentially introducing software bugs, security vulnerabilities, or incorrect decisions. Over time, heavy dependence on these systems could also erode important human skills, particularly in coding, mathematics, and problem-solving.
Looking ahead, Isola believes future AI agents will need capabilities that extend beyond today’s language models. More advanced systems may require new architectures capable of understanding video, physical interactions, sensor data, and other continuous information. Whether this evolution comes from enhancing existing language models with additional tools or from entirely new AI designs remains an open research question. Ultimately, the future of agentic AI will depend not only on making systems more autonomous but also on ensuring they remain reliable, transparent, and aligned with human goals.
