Course 2 - Module 4 - Lesson 2
When a workflow needs an agent and when a simple automation is enough
Useful Agent Recipes
Use fixed steps when the route is known
A simple automation follows rules chosen in advance. If a form always contains a confirmed date and contact address, software can copy those fields into a reminder template. There may be no need for a model to choose the next action.
An agent is more relevant when the route depends on observations. Investigating why a report is missing may require checking a folder, inspecting a job status, and deciding which evidence to request next. That flexibility also creates more behavior to inspect and control.
Break the workflow into different kinds of work
Consider a workshop registration process. Checking whether the required fields exist is a fixed rule. Drafting a friendly response to an unusual question may benefit from AI. Sending the approved confirmation is another fixed step. One workflow can combine all three.
Do not give the drafting step authority to change capacity or promise a refund. It can suggest language while the application enforces the registration rules. Separating these responsibilities makes the workflow easier to test and reduces the number of choices left to the model.
Choose with a practical decision test
Ask whether the possible inputs and branches can be listed clearly. If they can, start with a template or ordinary automation. If the helper must interpret varied information and choose among tools, consider a tightly bounded agent with observable decisions.
Compare total effort, including setup, failures, and review. A flexible agent may be unnecessary for a task that happens twice a year. Trial a small sample manually before building anything, and keep a human route for cases the process cannot handle.
Tiny experiment
Split workshop registration into three steps. Label each fixed rule, AI drafting, or agent decision, and explain whether any step really needs an agent.
Confidence check
Where would predictable rules make your workflow easier to trust than allowing the model to choose freely?