← Back to journal

Does this actually need an agent?

Before choosing a framework, write down which decisions are already known and which ones depend on what the model finds.

A new editorial note for this website, based on the linked public source. The English version is a translation.

We call a lot of different things “agents.” That can make a simple architecture decision harder than it needs to be.

Anthropic’s Building effective agents makes a useful distinction: a workflow follows paths defined in code; an agent gives the model more control over the process and tool use.

Start with the decisions

For a message-routing task with clear categories and destinations, a defined workflow may be enough. A request that needs different searches and tool choices each time raises a different design question.

A small exercise

Write the task in five steps. Beside each step, note whether the decision is fixed or depends on information found along the way.

Then define a good result, an error you care about and a stopping point. These are useful questions even before you pick a model.

The goal is to solve the task with an amount of complexity you can explain and evaluate.

Read Building effective agents

Ali Rashedi