Applied LLM Systems: RAG, Agents & MCP / Agents and Tools

What Is an Agent?

Model vs. agent.

L3 · AdvancedFast-movingKnown~10 min

Reviewed by YuvarajPublished 23 Aug 2026

Most uses of a language model are a single exchange: one prompt in, one answer out. An agent is different. It pursues a goal over many steps, deciding as it goes, calling tools, reading the results, and choosing what to do next. The model is still the engine; the agent is the loop and the scaffolding built around it.

Model vs. agent

A plain model call is a function: text in, text out, no memory of the world and no way to act on it. An agent wraps that call in a system that can hold a goal, act, observe, and decide whether to continue.

Plain model call

One prompt in, one answer out
Only produces text
No memory between calls
Cannot fetch live data or act

Agent

Pursues a goal over many steps
Chooses actions, usually tool calls
Carries a running trace as memory
Observes results and adapts

The single most useful sentence to remember:

A model predicts. An agent acts, observes, and adapts, in a loop, until a goal is met.

The parts of an agent

Check your understanding

Answer from memory before revealing, retrieval practice is what builds durable recall.

1.
What distinguishes an agent from a single model call?

Ask the tutor

Ask about this lesson, or about anything in AI. Answers cite the lessons they draw on.

Finished this lesson?

Mark it complete to earn XP, keep your streak, and schedule a review.

Nearly every agent, however it is built, has four ingredients:

Goal, what "done" means for this run.
Tools, functions it can call to fetch data or change the world (search, a calculator, an API, code execution).
Memory, what it carries between steps: the running trace, and sometimes longer-term storage.
Loop, the control flow that repeats: decide, act, observe, reflect. The next lesson is entirely about this loop.

Why the loop matters

Because the model cannot know a tool's result in advance, an agent cannot plan everything up front. It proposes one action, the environment answers, and only then can it choose the next action. That feedback is what lets an agent handle tasks a single prompt cannot, multi-step research, using live data, or recovering when a step fails.

For instance, asked "is the office open on the next public holiday?", an agent might first call a calendar tool to find the date, read that result, then call an hours tool for that specific date, a second step it could not have formed until the first returned.

Capable is not infallible

An agent that can act can also act wrongly, call the wrong tool, loop without progress, or take a step that is hard to undo. Real systems add guardrails: step budgets, tool permissions, validation, and a human in the loop for consequential actions. "Autonomous" never means "unsupervised." We return to this in Workflows vs. Agents.