Harrison's In the Loop What is an AI agent?

Harrison Chase explains that an AI agent is a system that uses an LLM to decide control flow, but acknowledges that the term is used differently across the community. He proposes an agentic spectrum, comparing it to autonomous vehicle levels, ranging from simple routers to fully autonomous agents that build tools and remember them. The article emphasizes that higher agentic systems require durable execution, background runs, and the ability to observe and modify state during runtime.

Chase highlights the need for evaluation frameworks that can test intermediate steps and handle randomness, as well as monitoring tools that let developers drill into each step. He notes that frameworks like LangGraph and LangSmith provide the necessary abstractions for branching logic and cycles. The piece encourages developers to assess the level of agentic behavior in their systems to guide framework choice and tooling. It also points out that agentic systems are more complex to run and interact with, necessitating background jobs and durable execution. Finally, Chase suggests that a clear understanding of agentic capabilities can improve development efficiency and robustness.