Start with the simplest thing that works

Among the practical writing on systems built with language models, "Building effective agents", published by Anthropic in December 2024, is one of the pieces worth reading slowly. It does not sell a promise. It puts the words in order.

Two different things with the same name

The central distinction is this. In a workflow, the path is written in code: the model is called at planned steps, and the program decides what comes next. In an agent, the model decides the path, picks the tools and keeps going until it thinks it is done.

They look like variations on one theme, but they behave very differently: a workflow is predictable and cheap to test; an agent is flexible, but costs more, takes longer and fails in ways that are harder to anticipate.

The patterns, briefly

The text describes a few workflow patterns that cover a lot of ground: chaining prompts with a check in between; routing each request to the right handling; parallelizing independent parts; an orchestrator that splits work across others; and a loop where one model evaluates and improves another's output.

The advice that stays

The recommendation I agree with most is the least spectacular one: look for the simplest solution possible, and add complexity only when it is needed. Often a single well-thought-out call, with good context, is already enough. An agent earns its place when the problem is open-ended, when you cannot predict the steps in advance, and when you can trust the model's judgment within limits.

It is the same instinct as any good craft: the most sophisticated tool is not the best one, the one the problem asks for is. And it pays to measure, to know whether the extra complexity is actually earning its keep.