THE BASICS / AI AGENTS
An AI agent.
To do what, exactly?
Finding information, preparing a record, using a tool: these possibilities do not all call for the same architecture. The word “agent” is not enough to choose by.
Reviewed by Saphan · 7 September 2026
Three ways of working to tell apart
| Form | How it works | Fictional example |
|---|---|---|
| Conversational assistant | Prepares a reply to a request | Rewording a note |
| Workflow | Follows defined steps, with or without AI | Extracting fields, then preparing a record |
| Agent | The model can choose steps or tools within a framework | Looking up a reference, consulting its record, asking for clarification |
The distinction proposed by Anthropic between workflows and agents helps to understand this difference. Products sometimes mix several forms behind the same interface.
Choosing from a client request
Imagine a question about the availability of a part. If an exact reference is provided, reading the stock and a presentation rule may be enough. If the client describes a machine without a reference, the system may need to look for a match and ask for confirmation before going further.
The second flow benefits more from an ability to interpret. It also requires knowing when the match is too uncertain. A plausible but wrong reference does not help the client.
The result matters as much as autonomy
A useful system can stay at the preparation stage: a draft, a sorting, a proposed record. Further automation is decided once its errors and the consequences of a wrong action are known. Not everything needs to be made autonomous to save time.
Describe the accessible sources, the authorized tools, the expected output and the cases that must go back to the team. Then test the whole with a missing piece of information, a refused access and an ambiguous request. It is this concrete scope that makes one agent proposal comparable to another.
A demonstration to understand the connection
The Saphan demonstrator turns fictional tickets into a table using local rules. It calls no model and does not claim to be an agent. It shows the part that is often forgotten: keeping the source, flagging what is missing and preparing a format that is actually usable.