Chatbot, workflow, agent: the words get used interchangeably in industry, and the confusion is expensive. The real distinction is a single question: who owns the control flow?
Generates a reply to a message. Control flow: one turn, then back to the human.
Example: a support FAQ bot that answers “What’s your return policy?”
LLM calls arranged along a predefined code path: steps and routing written by a developer.
Example: classify ticket → look up account → draft reply from template.
The LLM owns the control flow: it decides the next action from observations and state, acts, and continues until a stop condition.
Example: “find and book the best venue”, which searches, calls, and comparisons happen is decided at runtime.
| Chatbot | Workflow | Agent | |
|---|---|---|---|
| Who owns control flow | Human (per turn) | Developer (in code) | The LLM, at runtime |
| Path across runs | One turn | Same steps every time | Changes with observations |
| Predictability | High | High: auditable | Lower: needs eval + guardrails |
| Best when | Answering is enough | Steps are known in advance | The path can’t be scripted |
Anthropic’s framing (this week’s reading): workflows use predefined code paths; agents direct their own process, and you should look for the “simplest thing that works.” (Anthropic, Building Effective Agents, 2024.)
Click each pattern. Deterministic, auditable, cheap, and often all you need. Knowing them keeps you from building an agent where a workflow would do.
Drag the dial. Value rises with autonomy, and so do risk and the oversight you owe it. The engineering discipline: pick the lowest autonomy that solves the problem.