Original practice question. The scoring rubric is our teaching guide. How we research and score answers
TL;DR: Chatbot describes a conversational interface; it can front an answer-only system, a fixed workflow or an agent. A workflow uses models inside a sequence of steps that engineers designed. An agent decides its own next step, choosing which tools to call and when to stop, within limits it is given. The difference that matters for a PM is who controls the path: your code or the model.
How to approach it
Give a one-line definition of each with a concrete example, then explain why the distinction matters for reliability, cost and testing. Use "agent" carefully: in 2026 the word is applied to almost everything, and a precise definition is what makes the design answers after it credible.
A strong answer
Use one task, a customer asking to change a flight, to show all three.
Answer-only chatbot in this example. The customer asks; the model replies with instructions: "go to Manage Booking, select your flight, choose a new date." It may use retrieval to find the right help article. It takes no action. The customer does the work.
Workflow. Engineers design the steps: identify the booking, look up change rules for that fare, show available flights, confirm the price difference, make the change. A model is used where input is messy (understanding "move me to the evening flight on the same day"), but the order of steps and the actions are fixed in code.
Agent. The model is given tools (look up booking, search flights, check fare rules, change booking, issue refund) and a goal. It decides which tools to call, in what order, how to respond to unexpected results, and when it is done. If no evening flight exists, it might search the next morning, check whether a refund is cheaper, and ask the customer to choose.
| Answer-only chat example | Workflow | Agent | |
|---|---|---|---|
| Takes actions | No | Yes, fixed ones | Yes, chosen by the model |
| Path decided by | Not applicable | Engineers | The model, within limits |
| Control flow | One answer path; text may vary | Predesigned paths; model steps may vary | Model chooses paths within enforced limits |
| Handles unexpected cases | By explaining | Only cases designed for | Can adapt |
| Ease of testing | Easy | Step by step | Needs path-level evaluation |
These are not house definitions. Anthropic's guide to building agents (December 2024) draws the same line: in a workflow, predefined code decides how the model and tools are chained; in an agent, the model directs its own process and chooses its own tools. OpenAI's practical guide to building agents (April 2025) is explicit about the first column: a simple chatbot, a single-turn model call or a sentiment classifier uses a model without letting it control the flow of work, and so is not an agent.
The trade-off: each step towards agency adds flexibility and removes predictability. Anthropic's guide states the cost plainly: "The autonomous nature of agents means higher costs, and the potential for compounding errors." Agents can handle situations nobody designed for, which is their value. They can also take paths nobody approved, loop, or call tools in the wrong order, and every extra step compounds the chance of error. So the useful default is the least autonomy that does the job: a workflow when the path is known, an agent only where the path varies in ways rules cannot capture.
Many real products mix the two. A workflow handles the standard change; an agent handles the messy exceptions, with tight limits on what it can do.
What interviewers probe next
"Is a chatbot with retrieval an agent?" No. Retrieval adds knowledge, not autonomy. It becomes agent-like when the model directs the next steps, such as deciding which evidence to search for and whether to continue. A fixed retrieval pipeline can still appear in a chat interface.
"What limits would you give the flight agent?" Which tools it can call, a cap on refund amounts, confirmation before any charge, and a step limit after which it hands over to a human.
How it's scored
Weak: Uses "agent" for any LLM feature, or defines agents only as "more advanced chatbots".
Hire: Defines all three correctly, centred on who controls the path and whether actions are taken.
Strong Hire: Illustrates with one task across all three, explains the reliability and testing trade-off, and recommends the least autonomy that does the job.
Common mistakes
Inferring the architecture from a chat interface or from retrieval alone.
Treating agents as strictly better than workflows.
Forgetting that agents need limits: tools, budgets and approval points.
See workflow vs agent.