Agent, chatbot, workflow, or script?
Lesson 2 of 5 in What Is an Agent? The Loop That Changes Everything.
Almost every AI product page in 2026 says “agent” somewhere. To cut through it, ask three questions of any system:
- Who decides the next step? A human (at design time or at the keyboard), or the model at runtime?
- Is the behaviour bounded or open-ended? Can you enumerate everything it might do, or only describe the goal?
- How do you test it? Assert exact outputs, or run it many times and evaluate the distribution?
Those three questions sort the world into four boxes — and the boxes have radically different engineering, cost, and risk profiles.
| System | Who decides the next step? | Bounded or open-ended? | How do you test it? |
|---|---|---|---|
Script / automation — cron job, thermostat, support macro | A human, at design time. Every branch is written down. | Fully bounded — its complete behaviour is enumerable. | Unit tests. Same input → same output, every time. |
Workflow — Zapier zap, CI pipeline, LLM-summarise-then-file | A human, at design time. May call a model inside a step, but the step order is frozen. | Bounded step graph; individual LLM steps have fuzzy outputs. | Test the pipeline deterministically; eval the LLM steps statistically. |
Chatbot — ChatGPT, a RAG support bot, Copilot autocomplete | The human in the conversation. The model answers; the person decides what to do with the answer and what to ask next. | Open-ended text, but zero side effects — words, not actions. | Evals on response quality; no action trace to verify. |
Agent — coding agent, email triage loop, research agent | The model, at runtime. Each tool result shapes the choice of the next step. | Open-ended actions — the path through the tools cannot be enumerated in advance. | Run it many times; judge traces and outcomes statistically. Exact-output assertions are impossible. |
Interactive sorting exercise: Sort these ten real systems. The test is always the same: who decides the next step, and does the decision have hands?
Tool: Is It an Agent? — Think you can spot the boundary? Judge a stack of real systems — including the deliberately ambiguous ones — in the Is It an Agent? simulator.
Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.