AIM-06-001-P · Primer — What Is an AI Agent

The idea in one sentence

An AI agent is a system that uses a language model to decide its own steps—choosing and using tools in a loop to accomplish a goal—rather than following a fixed script.

Why it exists

Think of a personal assistant who not only follows a checklist but decides on the fly which calls to make, which calendars to check, and which restaurants to query based on what it learns; this mirrors how a human would handle a last‑minute dinner plan. The everyday problem is that static workflows break when conditions change—like a restaurant being fully booked—so we need something that can adapt dynamically. By giving the model control over the next action, we solve the brittleness of hard‑coded sequences and enable more flexible, goal‑oriented interactions.

How it works

The process starts with a user prompt, such as "book me a table Friday." The language model first interprets the goal and decides the first tool to invoke—checking the user's calendar. It calls a calendar API, receives the available times, and then decides the next step: searching for restaurants that fit the time slot. It queries a restaurant database, selects a candidate, and then calls a booking tool to reserve a table. If the booking fails because the slot is taken, the agent loops back, picks a new restaurant or a different time, and retries, continuing until the goal is satisfied or a timeout is reached. This loop embodies the shift from a simple LLM call to CONTROL, where each step is chosen based on prior results, matching the definition and spectrum of autonomy described in the facts.

A single‑tool‑augmented call is the minimal form of autonomy, while a fully autonomous agent plans, acts, and self‑corrects across many steps. Workflows represent a middle ground: predefined multi‑step paths that lack the dynamic decision‑making of agents.

See it — model + tools + loop

flowchart LR
  A["LLM<br/>(reasoning)"] --> B["+ Tools<br/>(act)"]
  B --> C["+ Loop<br/>(iterate to a goal)"]
  C --> D["= Agent"]

The one gotcha

A common mistake is assuming more autonomy always yields better results; in reality, agents add cost, latency, and unpredictability, so the simplest solution—often a fixed workflow—should be used when it suffices.

The one-line takeaway

Use an AI agent only when you need the system to decide its own next actions, not just follow a preset script.

Where you'll meet it

Practitioners encounter agents in platforms that expose tool‑calling APIs, such as Anthropic's Claude with function calling or OpenAI's function‑call feature. In audits, you’ll see agents flagged when logs show loops of tool invocations driven by model decisions, contrasting with single‑step API calls in standard LLM usage.