A fixed workflow with one AI step, or an agent that picks its own next step in a loop
A workflow runs the same steps every time, even if one step is a model, while an agent chooses each step from the last result. Here is how to tell them apart and when each one earns its place.
The word agent gets attached to anything with a model in it. Most of those things are workflows. The difference is simple to state: in a workflow, you chose every step in advance. In an agent, the model chooses the next step from what the last one returned.
A workflow is a fixed sequence. A form is submitted, the data is cleaned, a record is written to the CRM, a message goes to Slack. You can add a model to one step: step three drafts a two-line brief about the lead. The model writes text, but it does not decide what happens next. Step four runs whether the brief is good, bad or empty.
This is what Zapier, Make and n8n all do when you drop an AI module into a Zap, scenario or workflow. The model is fenced: one input, one output, parsed and checked before anything downstream uses it. If the model fails, the lead still reaches its owner. A good workflow with an AI step is boring, which is what you want for lead intake.
An agent runs a loop. It is given a goal, a set of tools (search the CRM, look up a company, send a quote, ask a human) and a system prompt. On each turn the model reads what happened so far, picks a tool and its arguments, sees the result, and decides again. It stops when it judges the goal met or hits a limit you set.
That loop makes an agent useful when the path is not known in advance. It also makes it risky. A workflow fails in a predictable spot. An agent can take a wrong turn on step two and spend eight more steps making it worse. All three automation vendors ship agent features: Zapier Agents with their own activity quota, Make AI Agents on all plans, and n8n's AI Agent node with human approval for tool calls (checked November 2026).
Here is the test the AI Agents Fundamentals course uses. Imagine an input you did not plan for: a lead form with a joke company name, or a document in an unexpected language. A fixed workflow does what it always does, possibly writing garbage into the CRM. A chatbot answers and stops. An agent looks the company up, asks a person, or tries another tool. If your system cannot change course at a surprise, it is a workflow, however many models it holds.
Pick a workflow with one AI step when the path is the same every time and only one step needs judgment. Lead intake, invoice coding, support ticket tagging and meeting summaries all fit.
Pick an agent when the steps depend on what you find: researching a prospect across five sources, triaging a bug report that could belong to three teams, reconciling a mismatch that needs several lookups. Even then, a person should approve any step it cannot undo.
Do not build an agent because it sounds better. A small ops team running one lead intake does not need a loop. A marketer who wants a brief drafted per lead needs one fenced model step with a daily cap.
A workflow is judged by its record: open the CRM contact the run created and check every field. The AI Automation with Zapier, Make and n8n course spends a chapter on that gap, then has you price, prompt, parse and cap one model step inside the same Tally-to-HubSpot intake, built on all three tools.
An agent cannot be judged by one demo, because the same question can get two answers. The agents course has you write a system prompt in five parts, turn a Python function into a tool with a description and argument schema, then run the agent against a fixed test set several times and count the correct runs. When one fails, you read the trace forward to the first wrong step.
If you copy data between apps and want a model to draft or classify inside that flow, take the automation course; it needs only short formulas and a few small code steps. If you write Python and have called a model from code, take the agents course.