Most things called agents should be one good prompt

A naming problem with a budget attached

Agent has become the default word for any AI feature, which is unhelpful, because the architecture it describes is genuinely different from a single well-scoped call and genuinely more expensive to build and operate.

Most systems being described as agents are one step: classify this, extract that, draft this. Built directly, they are cheap, fast, and easy to evaluate. Built as agents, they acquire orchestration, retries, and state for no gain.

The distinction is worth defending because the wrong choice shows up as unreliability, not just cost.

When multiple agents genuinely help

The case for separate agents is separation of concerns. A research step that gathers material, a writing step that uses it, and a checking step that evaluates the result are genuinely different jobs, and a single prompt asked to do all three does each of them worse.

That is not really about autonomy. It is the same reason you do not ask one reviewer to check accuracy and formatting simultaneously: narrow tasks produce reliable output.

The other real case is a task that branches unpredictably and has to take actions across systems based on what it finds. That is a genuine agent, and it is rarer than the marketing suggests.

What complexity costs you

Every additional step is a place to fail, and multi-step systems fail in ways that are harder to diagnose because the error surfaces several steps from its cause.

They cost more to run, since each step is a call. They are slower. And they are harder to evaluate, because a poor final output could originate anywhere in the chain.

None of that is disqualifying when the complexity is buying something. It is simply not free, and it is often bought without noticing.

A reasonable default

Start with the simplest thing that could work, which is usually one well-scoped call with good grounding. Measure it against real examples.

Add a step only when you can name the specific failure it fixes. Adding a quality check because output quality is inconsistent is a good reason. Adding orchestration because agents are the current architecture is not.

Systems built this way tend to end up with two or three deliberate steps rather than a graph nobody can reason about.