What is an AI agent?

The shortest definition that actually holds up: a chatbot answers your questions; an agent works to get things done for you.

An AI agent is software that, given a goal, can perceive its context, plan steps, use tools, remember what happened, and carry a multi-step task to completion with limited human steering. AWS puts it simply: humans set the goals, but the agent "independently chooses the best actions it needs to perform to achieve those goals."

How an agent relates to the model

The large language model is the "brain." Google Cloud describes the LLM as the brain of an agent, processing and generating language, while other parts let it reason and act. AWS frames the model as the reasoning engine that turns prompts into actions, decisions, or queries to tools and memory.

So an agent is the model plus the parts that let it act: perception (taking in context), planning, tool use, and memory. The model knows things; the agent gets things done.

What this looks like in practice

Anthropic, whose tools Berth manages, defines an agent as a system where the model "dynamically directs its own processes and tool usage." In practice that is an LLM using tools in a loop: it acts, reads real feedback from the environment, assesses progress, and repeats until the goal is met.