There are three levels of automation, and each one is more capable and harder to trust than the last. Most hiring problems are solved at the middle level. Most people reach for the top one, because that is the one being sold.
The three levels
A prompt is one instruction. You give it something, it gives you something back, you decide what happens next. You are in charge of every step.
A workflow is a fixed sequence you set up in advance. Read this CV, then score it, then if the score is high enough, draft a message. The software does the steps. You decided the order, and it never changes.
An agent chooses its own steps. You give it a goal and some tools, and it works out what to do, checks the result, and keeps going until it thinks it is finished or something stops it.
| Prompt | Workflow | Agent | |
|---|---|---|---|
| Who decides the next step | You | You, in advance | The software, as it runs |
| Same input, same path? | Yes | Yes | Not necessarily |
| How you fix it | Read the output | Find the step that broke | Read a trace and work it out |
| Cost | Exact | Predictable | Varies |
| Places it can go wrong | One | As many as there are steps | No fixed limit |
Only go up a level when you have to
This is the whole lesson. Letting software decide things for itself is not a feature you want for its own sake. It is a cost you accept when the problem genuinely needs it.
Hiring suits the middle level unusually well, because you already know the process. You are not discovering how to screen a candidate — you do it the same way every time. What you want from automation is doing a known job reliably and quickly, and that is exactly what a workflow gives you and an agent gives up.
Where agents genuinely earn it
- Open-ended research, where what you do next depends on what you just found — market research, company research, tracing someone's public work across sites.
- Inputs you cannot predict, where you could not list the cases in advance.
- Tasks where success is easy to check automatically. Agents are much safer when you can tell without a human whether it worked.
Notice that none of those involve decisions about candidates or messages to candidates. That is not an accident. Agents fit where a wrong answer is cheap to throw away.
Why more freedom goes wrong faster than it helps
A workflow with six steps has six places to fail, and each fails in a way you thought about when you built it.
An agent with six tools has an unlimited number of possible paths. And the failures are usually not in the individual steps — they are in the order it chose.
Worse, the mistakes build on each other. A wrong conclusion at step two becomes the input to step three, and by step seven the output is confident, coherent and wrong, with a plausible-looking trail behind it. Nothing crashes.
This is the "makes things up" problem from Module 1, given a budget and a set of tools.
A sensible default
Build the workflow. If one step genuinely cannot be written down in advance, make just that step agentic, keep it tightly limited, and have a person check it. An agent inside a workflow is far easier to defend than a workflow inside an agent.
The question to ask before building anything
"If I had to write down every step this will take, could I?" If yes, write them down. That is your workflow — and you have just avoided most of the problems in the rest of this module.