Workflows vs agents: how do you decide?

The difference is who controls the path. In a workflow my code decides the sequence: for example, classify the ticket, retrieve the policy, draft a reply, check it against rules, send it for review. The model does specific jobs inside fixed steps. In an agent, the model decides what to do next, which tool to call and when it is finished.

I default to workflows when I can write the steps down, because they are predictable. Each step can be tested on its own, costs and latency are bounded, and it is easy to explain to compliance what happened. Agents make sense when the next step genuinely depends on what was found, such as investigating a production incident, where you cannot know in advance which logs or services matter.

Most real systems end up hybrid: a workflow that handles routing, permissions and final checks, with one bounded agentic step inside it, such as research with a step limit. I would start with the workflow version, measure where it fails, and add autonomy only where that failure is about needing flexible steps.