Learn › Agent Harness, the big picture › Stop 1
Stop 1 of 9 · Page
Three kinds of AI application
There are only three kinds, and one question tells them apart: who decides the next step?
One question sorts them
Ask of any AI application: who decides the next step? There are only three answers.
| Kind | Who decides the next step | An insurance example | Good at | Watch out for | Harness needed |
|---|---|---|---|---|---|
| Single call | Nobody. There is only one step. | Turn a 12-page surveyor's report into five lines | Cheap, fast and easy to test | It only knows what you put in the request | A little: check what goes in and what comes out |
| Workflow | Your code. The steps and their order are written in advance. | Read a claim form, pull out the fields, check the policy, send an acknowledgement | Predictable: every run takes the same path, so it is easy to audit | It fails on a case that does not fit the steps you wrote | Some: a check at each step |
| Agent | The model. It picks a tool, sees the result and picks again. | A customer describes an accident in their own words; the assistant works out what to look up and what to do | Handles cases nobody wrote a path for | It can take a wrong step, a costly step, or too many steps | A lot: limits, permissions, approvals and a record of every step |
The same model can sit in all three.The kind is not about how clever the model is. It is about how much of the route you hand over to it.
How to choose
- Can one request do the whole job? Then it is a single call. Stop here.
- Do you know the steps before you start? Then write them down as a workflow, and use the model inside the steps that need it.
- Do the steps depend on what turns up along the way? Only then is it an agent.
Choose the simplest kind that does the job. Each step up buys flexibility and costs you predictability, money and testing effort.
The same three kinds in other industries
| Industry | Single call | Workflow | Agent |
|---|---|---|---|
| Banking | Sort an incoming complaint email into the right team | Check identity documents in a fixed order of steps | Look into an unusual payment across several systems |
| Hotels | Translate a guest review | Booking confirmed, then a welcome message, then an upgrade offer | A concierge that books, changes and fixes a guest's plans |
| Retail | Write a product description from a specification sheet | Return request, then the policy check, then the shipping label | A shopping assistant that searches, compares and places the order |
Next: the agent column is where the value is, and where the trouble is. Stop 2 looks at agents that work with less and less of a person watching.