In his keynote speech at the Inaugural ADAM symposium, September 15, 2025, “The Advent of Agentic AI” Dr. Kartik, Vice President & Chief Scientist, Systems Engineering, VAST Data shared that, in his opinion, it is through the creation of agents that AI can truly have impact. “LLMs are limited because they cannot take action…they can search the internet for flights, but they can’t book the ticket” Kartik emphasized. He also shed light on some of the pressing issues facing large language models (LLMs), including bias, hallucinations, factual inaccuracy, and limitations in reasoning. To illustrate these challenges, he demonstrated prompt responses in TinyLlama, a lightweight open-source model designed to help researchers experiment with LLM behavior on a smaller scale.
In the highlight of the talk, he introduced a simple Python-based agent that could detect when a user prompt requires a mathematical calculation—advanced reasoning that cannot be done by the AI. Instead of relying on the language model to guess the answer, which returned nonsense values that the LLM presented with supreme confidence in their accuracy, the agent automatically solved the problem using Python’s math libraries and returned the correct result. The demonstration underscored a practical approach to combining LLMs with external tools to improve accuracy, reliability, and trust in AI outputs beyond the capabilities of the models themselves.
