Andrej Karpathy, a prominent AI researcher and former Tesla AI director, recently voiced concerns on social social media regarding the increasing "agentic" behavior of large language models (LLMs). Karpathy noted that LLMs are becoming "a little too agentic by default" due to optimization for long-horizon tasks, leading to prolonged reasoning times and extensive tool use, particularly in coding environments. This trend, he suggests, often exceeds the requirements of average user interactions.
Karpathy elaborated on the practical implications, stating, "For example in coding, the models now tend to reason for a fairly long time, they have an inclination to start listing and grepping files all across the entire repo, they do repeated web searchers, they over-analyze and over-think little rare edge cases even in code that is knowingly incomplete and under active development, and often come back ~minutes later even for simple queries." This behavior, while potentially suitable for complex, long-running tasks, proves less effective for "in the loop" iterative development or quick spot checks.
The AI expert frequently finds himself interjecting with commands such as "Stop, you're way overthinking this. Look at only this single file. Do not use any tools. Do not over-engineer," to curb the models' expansive problem-solving approaches. He emphasized a growing need for clearer communication of user intent, ranging from "just have a quick look" to "go off for 30 minutes, come back when absolutely certain," as the default behavior gravitates towards an "ultrathink" super agentic mode.
This observation aligns with broader industry discussions about the evolution of AI agents. While the market sees a significant shift towards autonomous AI agents capable of planning and executing multi-step tasks, there's also a recognition of the need for precise control and adaptability. Companies and researchers are exploring ways to balance agentic capabilities with user-defined constraints, ensuring LLMs remain versatile tools for a wide spectrum of applications, from highly autonomous workflows to simple, direct queries. The ability to manage the level of agentic autonomy is becoming a critical feature for future LLM development.