AI Learning Hub › Learn AI › Self-Paced Courses › Agents Without the Hype › What Makes Something an Agent?

Module 1 ยท Beginner

What Makes Something an Agent?

Workflows follow a path chosen in advance; agents let the model choose the next step. Tell them apart by asking who decides, rather than by reading the product label.

By the end

  • Distinguish a one-response AI interaction from a workflow
  • Distinguish a fixed workflow from an agent
  • Identify who or what decides the next step
  • Explain why tools alone do not make a system an agent
  • Recognize that agent behavior can still operate within human-defined boundaries

Start with one question

There is one question that does more work than any definition:

Who decides what happens next?

If you ask a question and receive one response, that may simply be AI chat. You asked; it answered; the exchange is over.

If software follows a sequence selected in advance, it may be a workflow — even if AI appears inside some of those steps. Someone decided the order before you arrived, and the system runs it.

If the model can look at the goal and the results so far and make meaningful choices about what step to take next, the system is behaving more like an agent. What matters is how the path gets chosen, not how many steps it has.

Three words to learn

Goal

The outcome the user wants, without necessarily specifying every step required to reach it.

Workflow

A sequence whose path is substantially defined in advance.

Agent

An AI system in which the model makes meaningful choices about how to pursue a goal, often using tools and results from earlier steps to decide what to do next.

These are working definitions, not legislated ones. They are the distinction this course uses, and they line up with how Anthropic and OpenAI describe the difference in their own engineering guidance. They are not a standard that every vendor has agreed to follow.

Three systems, one question

Below are three fictional systems. For each one, decide whether you would call it a response, a workflow, or an agent.

Read the description, commit to an answer in your head, and only then open the reveal. Guessing first is the exercise; reading the answer first is not.

Scenario A

You paste a paragraph into an AI assistant and ask it to make the paragraph easier to understand. The AI returns a revised paragraph. The interaction ends.

Response, workflow, or agent?

Reveal the answer

Response.

One model response may involve sophisticated reasoning internally, but there is no ongoing process in which the system chooses among external steps toward a larger goal.

Notice which facts turned out to be irrelevant. The rewrite may have been excellent. The model behind it may be an extremely capable one. Neither changes the answer: the system was asked once, and it answered once.

Scenario B

A document system always performs the same sequence: summarize the uploaded text, convert the summary to bullet points, then format those bullets as an email draft.

Response, workflow, or agent?

Reveal the answer

Workflow.

The system may use AI at every step. It may even use multiple tools. But the route — summarize, then bullets, then email — was chosen in advance, by a person, before any document arrived.

Run it a thousand times and it takes the same path a thousand times. The word always in the description is the giveaway.

More steps do not automatically make a system an agent.

Scenario C

You ask an AI system to identify useful materials for a research question. It can search a collection, inspect the results, decide whether another search is needed, compare what it found, and stop when it has enough information — or ask you for help if the available material is insufficient.

Response, workflow, or agent?

Reveal the answer

Agent.

The system has:

  • a goal;
  • available capabilities;
  • intermediate results;
  • model-directed choices about what to do next; and
  • a point at which it stops or asks the human.

Nobody wrote down in advance how many searches this would take, or which comparison mattered. The number of steps depends on what the system finds along the way.

Note what this does not say. The system is not unsupervised, and it is not independent of you: it still reports back, it still stops, and it can still hand the problem to you. It is an agent because the model is choosing the route, not because nobody is watching.

Fixed path vs. chosen path

Setting the two architectures side by side makes the difference concrete. These are observable, system-level steps — things you could watch happen — not a claim about the model’s inner reasoning.

Workflow

The route is substantially chosen ahead of time.

  1. summarize;
  2. extract;
  3. format;
  4. finish.

A person decided this order. The system follows it.

Agent

The goal is defined, but the model can make meaningful decisions about the route.

  1. inspect the goal;
  2. choose an available capability;
  3. observe the result;
  4. decide what to do next;
  5. stop or check with the human.

A person decided the goal and the boundaries. The model decides the order.

This is not about hidden reasoning. Some AI systems display their internal deliberation, and some do not. That is a separate topic. The steps above are external actions: a tool was called, a result came back, another tool was called. You could log them.

Tools do not make something an agent

This is the most common shortcut, and it is wrong.

The presence of a tool is not enough to determine the architecture. “It can search the web” and “it can read your files” describe reach. They say nothing about who is steering.

So the question worth asking about a tool is this:

Who decides whether to use the tool, which one to use, and what to do after seeing the result?

If a person wired that decision in advance, you are looking at a workflow with tools. If the model makes it while the task is running, you are looking at something more agent-like.

Already completed MCP Without the Mystery? Think of MCP as one way an AI host can gain capabilities. Agents are about how those capabilities may be selected and sequenced toward a goal. The two courses are independent, and neither requires the other.

Agent does not mean “no human”

An agent can still work within a narrow task, have only a few available tools, require approval before it acts, stop when it encounters uncertainty, and return control to the user.

All of those constraints can be true at once, and the system is still an agent, because the model is still choosing the route within them.

The useful question is not “Is it autonomous?” but “Which decisions has the system been allowed to make?”

An agent is not always the better design

Both Anthropic and OpenAI, in their own published guidance for people building these systems, recommend matching complexity to the task rather than reaching for an agent by default.

If the steps are known and predictable, a workflow can be easier to understand, easier to test, and easier to supervise. You can read the sequence. You can predict what it will do tomorrow. You can explain it to someone else.

Agent behavior becomes useful when the route cannot be completely specified in advance and the model needs to adapt based on what it encounters — when the right second step genuinely depends on what the first step turned up.

Use the simplest system that can reliably do the job.

This is worth stating plainly, because the surrounding conversation often implies the opposite: that workflows are what we used to do and agents are what we do now. That framing is not how the people building these systems describe the choice. Neither one is a more advanced version of the other. They are different answers to the question of who plans the work, and the right answer depends on whether the plan can be written down ahead of time.

Try one yourself

One more, without a label attached:

Scenario D

An AI receives a set of documents and a checklist. It must check every document against the same five criteria in the same order, then produce a table.

Would you describe this as a workflow or an agent? What fact matters most?

Reveal the discussion

Workflow, assuming the path truly is fixed.

The important fact is not that AI analyzes the documents; it is that the sequence is predetermined rather than dynamically selected. Every document gets the same five criteria in the same order, and the system has no say in that.

Notice how much sophisticated AI work is happening here. Reading a document and judging it against a criterion is hard. None of that difficulty makes the system an agent, because none of it involves choosing what to do next.

Now change one thing. If the system were allowed to decide which documents required additional investigation, choose new tools, or change its approach based on intermediate findings, it would become more agent-like under the working distinction used in this course.

That is the fact that matters most: not how clever the analysis is, but whether the system is permitted to alter its own route.

What to notice

Agent is a description of how control is organized, not a compliment for a sophisticated AI system.

A simple workflow may be exactly what a task needs. Calling it an agent would not improve it, and calling an agent a workflow would not make it safer.

An agent becomes interesting — and worth examining carefully — when the model has meaningful choices about how to pursue a goal. That is when the question of what it is allowed to reach and change starts to matter, and that is where the rest of this course goes.

Before you continue

You are ready for Module 2 if you can: