The Sequence Opinion #552: Seriously, What is an Agent?
I typically use this section to discuss original ideas about AI and not worry too much about industry buzzwords but wanted to make a bit of an exception today. In a recent discussion with a friend that works in a major AI lab, we came up with about a dozen definitions for what we call AI agents today. So much so, that I thought it merited an entire installment of this newsletter. Read on, I think you will be surprise ;).
You ask a room of experts what an "AI agent" is, and you may get wildly different answers. In recent years, agent has become an umbrella term applied to everything from simple chatbots to ambitious autonomous systems. This breadth of meaning reflects an ongoing debate in the tech community about what truly defines an agent. Is it the ability to plan and act? Is it autonomy from human guidance? Does using a large language model make something an agent, or is an agent more than just a fancy function call? This essay explores the diverse conceptualizations of AI agents. We will examine eight perspectives on what an agent can be, from structured planners to tool-wielding automatons, and discuss how these facets are converging. In closing, we consider whether the boundary between the AI model and the agent is beginning to erode, heralding a future where agentic capabilities might be built directly into the models themselves.
Join Me for a Chat About AI Evals and Benchmarks:
