Analysis
Lexicon
Lexicon · III · Orchestration
Agent
Also: AI agent
An AI model allowed to chain actions toward a goal (act, observe the result, correct) instead of answering once and stopping.
The analogy
The assistant is a switchboard operator: they answer each call, precisely, then hang up. The agent is the temp worker you hand a mission to: "take care of this file." They open the file, hit an unexpected problem, make a call, fix it, and come back when it is done, or when they are stuck. What you handed them is no longer a question. It is an outcome to reach.
A concrete example
Give an AI the goal "find out why invoice 4021 was rejected." An assistant would explain the possible causes of invoice rejection. An agent opens the billing tool, reads the error logs, forms a hypothesis, checks it, and comes back with the actual cause, without being told any of those steps. Every round of "act, observe, correct" is one turn of the loop; that loop is what makes it an agent.
What it is NOT
It is not an upgraded chatbot: the leap is not in answer quality but in moving from answering to doing. And it is not magical autonomy: an agent is exactly as reliable as what its harness allows it to do, and as the supervision planned around it. An organization deploying agents without deciding those limits has not deployed AI: it has hired without a job description.
Often confused with
The distinction that changes the conversation.
- Harness
- An agent only exists once it is equipped with a harness: the harness gives it its tools and its limits. Talking about an agent without its harness is talking about a driver without saying what vehicle they drive.
- Assistant
- The assistant answers when asked, one question at a time, and you chain the steps yourself. The agent receives a goal and chains the steps itself. The difference is not power: it is who holds the loop.