Field Note · Agentic AI · August 2026

What is an AI agent — and what it isn't.

An AI agent is a system that uses a language model to plan and execute multi-step tasks — calling tools, checking results, and escalating to humans at defined checkpoints. The one-line test: an assistant augments a task you drive; an agent owns a task within guardrails you set.

The note

Most confusion about agents comes from stretching the word. A chatbot that answers questions is an assistant: the human decides what happens next after every turn. A workflow that follows a fixed script is automation: nothing decides anything. An agent sits between them — it is given an objective, plans its own steps, uses tools against real systems, and makes bounded decisions along the way.

"Bounded" is the operative word. In production, the boundary is engineered, not hoped for: explicit contracts for what the agent may, must not, and must-ask-before doing; human checkpoints on consequential actions; and evaluation that proves the boundaries hold. That's the architecture I describe in the AI Agent Production Stack — and it's why "how do we test this?" is the first question worth asking about any agent, covered in the agent evaluation guide.

A useful heuristic from teaching this at ASE Bucharest: if removing the human from the loop changes nothing, you built automation; if removing the model from the loop changes nothing, you built a form. An agent is what you have when both the model's judgment and the human's oversight are load-bearing.