When to use an agent (and when not to)

Most workflows don't need an agent. They need a well-designed pipeline with clear inputs and outputs. The distinction matters because agents introduce non-determinism, and non-determinism has a cost — in debugging time, in trust, and in the guardrails you now have to build.

I reach for an agent when a task genuinely requires judgment calls across a variable number of steps, where the path isn't known ahead of time. Claims triage is a good example: the agent might need to check policy data, flag a missing document, and re-check after a follow-up — the sequence isn't fixed.

If your workflow is 'do A, then B, then C', it's a pipeline, not an agent. Ship the pipeline. It'll be cheaper, faster, and easier to reason about — and you can always add an agent at the one step that actually needs it.

The teams that get the most value from agents are the ones who resisted using them everywhere.