MIT researcher explains where agentic AI works and where it falls short

Agentic AI refers to systems that take actions in the world, from booking flights online to manipulating objects with robots. Phillip Isola, an MIT associate professor in electrical engineering and computer science and CSAIL member, describes today’s agents as generative AI models wrapped with tools, memory, and application-specific interfaces that let them interact with websites, software, or physical systems.

Adoption has accelerated: a November 2025 report by MIT Sloan School of Management and Boston Consulting Group found that 35 percent of surveyed businesses had already deployed AI agents, while another 44 percent planned to implement agentic AI soon. Isola points to coding agents as one of the strongest current use cases because they can test solutions, receive feedback, and iterate until they find a working approach.

The main technical barrier is limited training data for real-world actions, such as knowing where to click, how to recover from errors, or how to negotiate a transaction. Risks include weak verification, leaked private data, software bugs, vague human instructions, and de-skilling as people rely on agents before the technology is ready to fully automate complex tasks.

Future agents may need models that can handle video, physical forces, radar scans, time series, and other data types beyond text. A major open question is whether more capable systems will emerge from language models equipped with sensors and tools, or from architectures built differently from the ground up.