Recommended resources
The freely available resources this lesson will direct you to.
Hello! This course starts with a question that will guide everything you build later: what makes an AI system an agent? In this first module, you will learn to recognize the difference between a system that replies, one that follows a prescribed procedure, and one that chooses how to pursue a goal.
Today, we will use concrete examples rather than treat “agent” as a label to memorize. By the end, you should be able to look at an AI application and ask who decides what happens next. That question is more reliable than asking whether the application sounds intelligent.
A chatbot describes the interaction, not the whole system
Imagine asking an AI chatbot, “What are three questions I could ask when researching a historical event?” It responds with suggestions. You might follow up, and it responds again. The central interaction is a conversation: you supply a message, and the chatbot supplies a reply.
A chatbot can use a language model to handle open-ended questions, but not every chatbot does. Some follow scripted rules. Equally important, a conversational interface does not tell you everything about what happens behind it. A chatbot could simply generate an answer, or it could sit in front of a more elaborate system.
Read these two short comparisons from IBM to establish what the word chatbot does—and does not—tell us.
IBM distinguishes scripted chatbots from AI chatbots, then compares chatbots with agents. Read it to see why a chat window alone cannot establish whether a system is an agent.
Under “Traditional chatbots compared to AI chatbots,” begin with the bullet describing chatbots that follow rules or use AI; stop after the bullet about AI chatbots receiving conversational responses. Then find “AI chatbots compared to AI agents.” Read the comparison, from its opening sentence through the two bullets describing typical chatbot and agent behavior. Focus on the difference between conversing with someone and carrying out work toward a goal.
For this course, picture the simplest chatbot as one that answers the current request. If you ask it for a research outline, it gives you an outline. It has not necessarily searched for sources, checked its suggestions, or decided to take another step after replying. Those actions would require additional capabilities and instructions.
This is not a claim that chatbots are incapable of doing more. A system can have a chatbot interface and operate as an agent underneath. “Chatbot” tells you how you communicate with it; “agent” tells you something about how it controls its work.
A fixed workflow can do several things without being an agent
Now imagine an application that prepares a short research briefing. Its developer specifies this procedure:
- Use a language model to draft an outline.
- Check whether the outline has three sections.
- If it passes, use another model call to write the briefing. If it fails, stop and report the problem.
This is a fixed workflow. Its output may vary, and a language model may write both the outline and the briefing. Yet the developer has already specified the available stages, their order, and what the check does. The model is contributing content within that structure; it is not choosing an open-ended route through the task.
The image below depicts this kind of workflow. The gate makes a conditional choice, but the pass and fail destinations are already defined. A branch does not, by itself, make a system an agent.
This distinction can feel surprising because a workflow may look quite capable. It might search a database, call several AI models, or run each morning without anyone pressing a button. Tool use, multiple steps, and automatic execution do not by themselves make it an agent. The question is whether the model is permitted to decide which actions to take next in response to what it finds, rather than merely supplying results to a path specified in advance.
Watch the following short portions of Jeff Su’s explanation. The calendar example makes the limitation of a fixed route particularly clear.
Watch “AI Agents, Clearly Explained” by Jeff Su for a visual contrast between a basic language-model reply, a fixed workflow, and an agent.
Watch basic replies for the prompt-and-response example. Then watch fixed routes for the calendar and weather example: notice that adding steps does not change who designed the route. Finally, watch the comparison for the three-way visual summary. You can skip the intervening material for this lesson.
There is a useful nuance here. A workflow can contain branches and even repeated steps; it need not be one straight line. What makes it fixed is that its controlling procedure is predefined. A model might classify a customer question as “billing” or “technical,” but if the program sends each category to a predetermined process, the model is selecting among routes that the developer laid out.
An agent chooses how to pursue a goal
Suppose you instead give a research assistant the goal: “Find out whether these two accounts of an event agree, and explain any disagreement using sources.” You have not told it which source to inspect first, how many sources will be needed, or what to do if an account omits a crucial detail.
In an agent, the language model has a decision-making role during the work. It might choose to look up one account, inspect what comes back, search for a second source, and decide whether it has enough information to answer. If a source is unhelpful, it can choose a different permitted action. Its route depends on what it encounters.
Contrast three possible responses to that same request:
| System | What it might do | Who determines the route? |
|---|---|---|
| Simple chatbot | Offer a comparison based on the text supplied in the conversation | You provide the material; it replies |
| Fixed workflow | Always retrieve two specified sources, summarize each, then compare the summaries | The developer specifies the stages |
| Agent | Choose which permitted sources to inspect and whether another search is needed before answering | The model makes bounded decisions during execution |
“Bounded” matters. An agent need not have unlimited freedom. A developer can restrict which tools it may use, how many steps it may take, and when it must ask for approval. Within those boundaries, the model still decides something consequential about the next action. We will study the mechanics of those decisions in the next lesson.
Read the focused sections below to connect this example to a more precise architectural distinction. The same article calls workflows and agents agentic systems in a broad sense, but separates them by who controls the process.
Anthropic’s “Building Effective AI Agents” gives the core distinction, a prompt-chaining example like the gate above, and a description of agents making decisions during execution.
Start in “What are agents?” and read the distinction in the two bullets defining workflows and agents. In “When (and when not) to use agents,” read the paragraph beginning “When more complexity is warranted” and ending “usually enough”; notice the trade-off between predictability and flexibility. Next, under “Building blocks, workflows, and agents,” find “Workflow: Prompt chaining.” Read prompt chaining through its examples, stopping before “Workflow: Routing.” Finally, under “Agents,” read agent operation, from the opening paragraph through the caution about testing and guardrails. Focus on how an agent uses what it finds during execution to inform its next decision.
An agent is therefore not simply “a better chatbot.” Nor is it defined by the number of model calls. A single reply can be useful; a fixed workflow can be elaborate; and an agent can be small. The defining difference is model-directed choice of actions in pursuit of a goal.
A reliable way to tell them apart
When an application is described as an “AI agent,” ignore the name for a moment and inspect one run of the application.
First ask, is it mainly replying to the user’s current message? If so, chatbot describes its visible behavior. Then ask, does it carry out a multistep task? If it does, examine who decides the steps. If the program specifies the route—including its branches—the underlying process is a workflow. If the model chooses among permitted actions, uses the results to reconsider what is needed, and determines how to continue toward the goal, it has an agent’s role.
Consider a daily news digest. One version always collects articles from the same feed, summarizes the first five, and formats an email. That is a workflow even if it runs unattended. Another is given a topic and permission to search: it chooses queries, judges whether the results cover the topic, and may investigate a missing perspective before drafting. That is agent-like behavior. Both could deliver their result through a chat window, email, or an app; the delivery format is not the deciding factor.
Real systems can combine these designs. An agent might call a fixed workflow to format its final report. A workflow might hand one difficult subtask to an agent, then resume its predefined stages. Rather than force the entire application into one box, identify which part follows a preset route and which part, if any, lets the model direct the work.
Finally, flexibility has a cost. When a model chooses actions, it may choose poorly, take unnecessary steps, or misunderstand a result. A predictable task often needs only a simple reply or a fixed workflow. The point of recognizing an agent is not to prefer one automatically; it is to understand what additional decisions you are entrusting to the model.
Takeaways
A chatbot is characterized by conversation. A fixed workflow follows developer-defined routes, even when those routes contain AI calls, tools, or branches. An agent pursues a goal by making bounded decisions about its actions as the task unfolds. To distinguish them, ask: who decides what happens next?
Next, we will look more closely at what an agent needs to make those decisions: a goal, actions it can take, and information it receives from the results.