Not everything needs to be an agent

A lot of AI work currently fails the same way: a clear procedure is wrapped in an agent because agents are the current shape of ambition.

If you already know the steps, you do not need a system that rediscovers them at runtime. You need a workflow. Workflows are cheaper, easier to evaluate, and honest about failure.

A simple rule

• If the path is known, use a workflow.
• If the next step depends on what the model discovers, consider an agent.

Agents earn their complexity when the environment is genuinely open: tools, incomplete information, a goal that cannot be expressed as a script. They do not earn it because a demo looks more autonomous.

Cost is architectural

An agent that loops, calls tools, and re-plans is a distributed system with a non-deterministic control plane. That can be the right design. It is rarely free. Latency, eval burden, and operational surprise arrive together.

The goal is not to make every product autonomous. The goal is the simplest architecture that reliably does the work.

What we do instead

We start with the job. We write the steps. We add a model where judgment is required. We add an agent only when the path cannot be written down in advance.

That is less fashionable. It ships.