MIT Associate Professor Phillip Isola explains agentic AI in a Q&A format, covering how it differs from generative AI, how agents are built on top of foundation models like Claude, and where they work best today. Coding agents are highlighted as the most mature application. Key risks discussed include insufficient human verification of agent outputs, vibe coding leading to bugs and data leaks, and de-skilling as humans offload cognitive tasks. Looking ahead, Isola raises the open question of whether future AI agents will simply be LLMs with more sensors and tools, or require fundamentally new architectures to handle continuous, high-dimensional, and physical-world data.
228 Impressions