Agent or Not?
You have now learned what makes a system an AI agent: it takes autonomous action in the world using tools, guided by a goal, across multiple steps. You know how the perceive-reason-act loop works, what tools enable, why autonomy matters, where agents shine, and where they struggle. Now it is time to test that understanding in the most important way — by applying it to real-world examples and defending your reasoning. Classification with justification is how expertise becomes usable.
When classifying any system, ask three questions: (1) Does it take actions beyond producing text? (2) Does it make autonomous decisions about what to do next? (3) Does it use tools to affect something outside the conversation? If all three are yes, it is an agent. If the answer to all three is no, it is a chatbot or something simpler. Mixed answers mean it is somewhere on the spectrum.
Warm-Up: Clear Cases
Before the main activity, make sure you are comfortable with the clearest cases on both ends of the spectrum. Definitely a chatbot: a customer service bot on a website that reads your question and types a helpful reply from a knowledge base. It produces text. It does not take actions. It does not call tools. It does not remember your name between visits. Definitely an agent: a coding assistant that reads your codebase, identifies a bug, writes a fix, runs the test suite to confirm the fix works, and opens a pull request — all without you directing each step. It uses file-read, code-execution, and version-control tools. It makes autonomous decisions across many steps. It produces real outcomes beyond text.
The Classification Activity
Agent Classification Challenge
- For each system described below, do three things: (1) Classify it as agent, not-agent (chatbot or simple automation), or on-the-spectrum (has some agent qualities but not all). (2) List the specific evidence from the description that supports your classification. (3) Write one sentence explaining what capability it would need to gain (or lose) to move fully to the other category.
- System A: A voice assistant on a smart speaker that answers questions, plays music when asked, and can set a timer — but cannot learn your preferences over time or act without a direct voice command each time.
- System B: A travel booking system that, when given a destination and budget, searches multiple flight and hotel databases, finds the best combination within budget, fills in your saved traveler profile, submits the booking, and emails you the confirmation — without asking for step-by-step approval.
- System C: A tutoring chatbot that explains math concepts, generates practice problems, and grades your answers — but all within a single conversation window, with no ability to update your grade record, send your teacher a progress report, or remember anything about you next session.
- System D: A social media manager that monitors your brand's mentions every 15 minutes, drafts a reply for any mention with a negative sentiment score below a threshold, and sends you the draft for approval before posting it.
- System E: A security monitoring tool that scans your network traffic continuously, automatically blocks any IP address that exceeds 1,000 failed login attempts in a minute, and logs the incident to a database — all with no human in the loop.
- System F: A homework helper that a student can chat with, and which remembers previous conversations within the same school year, can look up topics in a connected reference library, but cannot send messages or modify any files outside the chat window.
Discuss: The Gray Zone
You probably found that some systems were hard to classify. System D, for example, monitors continuously (agent-like), drafts autonomously (agent-like), but holds for human approval before posting (not fully autonomous). System F has persistent memory and tool access to a library, but cannot take actions outside the chat — it is more than a basic chatbot but less than a full agent. This gray zone is real and important. As AI systems grow more capable, many will live in this in-between space. The right question is not always agent or not-agent but rather: how much autonomy does this system have, over what actions, and how well is that autonomy supervised? Those are the questions that matter for safety and design.
Match each system feature to the correct classification it most supports.
Terms
Definitions
Drag terms onto their definitions, or click a term then click a definition to match.
A system reads incoming customer complaints, categorizes each one as billing, technical, or general, and automatically routes it to the right department's inbox — with no human review. Is this an agent?
What is the best follow-up question to ask when a system falls into the gray zone between chatbot and full agent?