A fixed workflow follows the same steps every time. An agent decides its own steps. That one difference should drive what you buy, what it costs and how much checking it needs. A job you can write down step by step is usually better as a workflow, whatever the product is called, and the four-question test below will help you tell which is which. If you want the service pages first, they are custom AI agents and workflow automation.
What is the difference between a workflow and an agent?
A workflow runs steps you fixed in advance, while an agent gets a goal and tools and picks its own steps. A workflow says: when this happens, do step one, then step two, then step three. A new lead fills out a form, the workflow adds it to your CRM, sends a confirmation and posts a note in Slack. It can include a step where an AI model reads text or drafts a reply, but you chose the order in advance. n8n says it plainly: the more fixed the sequence, the better a workflow fits, even if one of the steps is a model making a decision.
An agent gets a goal and a set of tools, then works out the steps on its own. n8n describes its new Agents feature as something you describe in plain language, give a model and the tools it may use, and it works out the steps itself. GrossiWeb's agent page defines the same idea as software that reads inputs, decides which tools to call, acts in other systems and handles errors without a person directing each step. The cost of that freedom is predictability: with a workflow you can read the steps in advance, and with an agent you review what it did afterward.
A job you can write down is a workflow, whatever it is called
If you can write the steps on a whiteboard before the job starts, you need a workflow, not an agent, whatever the product is called. n8n's own example of a workflow job is a lead that comes in, gets enriched, scored and routed.
Fixed steps are cheaper to build and test, run the same steps every day, and are easier to debug. GrossiWeb's agent page says as much: simpler pattern-based tasks are often cheaper as workflow automation. Our earlier post on saving 10 hours a week with n8n shows that shape in practice: set steps, a model for the language work, and a person approving before anything sends. Whatever a product is called, ask the vendor who decides the steps, you or the software. If the answer is you, it is a workflow. For a wider check on whether a tool is worth buying at all, read why most AI tools waste your time.
When an agent is the better choice
An agent earns its place when the input is different every time and the next step depends on the last answer. n8n gives the example of a teammate asking in Slack why a customer's usage dropped last month. Answering takes a few rounds: pull the account, ask which product line, check support history, come back with a summary. Nobody can lay that process out in advance. n8n puts a support email or a new GitHub issue in the same bucket: requests where you know the domain but not the task.
Even then, the strongest design is not an agent running loose. In n8n's support example, the agent decides when each of three workflows runs, but what happens inside a workflow stays fixed, step by step. One of those workflows only adds a note to an account, so the agent never holds the CRM credential. That pattern is the practical answer: start with a workflow, then add an agent only where a step needs judgment. Zapier's Next Gen Zaps, still in early access, add a monitoring agent that watches runs and can apply or suggest a fix, and n8n labels its agents as still in preview. Test either before you bet a process on it.
The four-question decision test
Ask four questions about the job: is the input predictable, what does a mistake cost, who checks the output, and how is usage billed. Predictable input, expensive mistakes and a tight budget point to a workflow. Messy input, cheap mistakes and a reviewer point to an agent, ideally one that calls your workflows as tools.
1. Is the input predictable?
If the same form or email format arrives every time, a workflow handles it. If the input is free-form and the next step depends on what the person wrote, an agent may be justified. Take ten recent examples of the job: if the same steps cover all ten, you have your answer.
2. What does a mistake cost?
A wrong internal summary costs a few minutes. A wrong email to a customer or a wrong change to a billing record costs money and trust. The higher the cost, the more you want fixed steps and a narrow list of allowed actions. n8n's advice for anything sensitive is to start where the blast radius is small: scoped tools, a test channel and approvals on any action that writes to a system of record.
3. Who checks the output?
Name a person before you build anything. A workflow produces the same kind of output each time, so spot checks are quick. An agent can take a different path each run, so someone has to read its logs. n8n lets you mark a tool as sensitive so the agent pauses for a person to approve or reject before using it.
4. How is usage billed?
Billing differs by tool, so check it before you scale. n8n counts executions, and an execution is a single run of the entire workflow, however many steps it has. Under n8n Agents, one turn with an agent is one execution, and agents share the same quota as workflows. On October 2, 2026, n8n's pricing page listed Starter at 20 euros a month billed annually for 2,500 executions and Pro at 50 euros a month billed annually for 10,000.
Zapier counts tasks, and a task is any successful action that runs. A run with five action steps uses up to five tasks: Filter, Paths, Formatter and Delay steps are free, while AI by Zapier steps use tasks at a rate set by the model tier you pick. The model can be a further cost: an n8n agent runs on a model you hold credentials for, billed by that vendor, or on n8n's Gateway credits. An agent that takes many steps per job can therefore cost more than a fixed workflow that calls a model once. A separate post of ours compares the three tools' pricing in depth.
What to do this week
You need a clear list this week, not a purchase decision. Sort your repetitive work with the steps below before you talk to any vendor, including us.
- List three recurring jobs that eat time, and write the steps for each on one page.
- Mark every job where you cannot write the steps in advance. Those are the only agent candidates.
- Build the fixed jobs as workflows first. If a step needs a model to read messy text, put the model inside that one step.
- For anything that sends a message or changes a customer record, keep a person approving the output until you have reviewed enough results to trust it.
- Look up how your chosen tool bills runs, then estimate a month of volume.
If you want help sorting the list, the automation and scale page explains how GrossiWeb approaches it. MCP, the connector standard that lets agents reach other tools, gets its own post.