workflow

In depth

What gets written first

A workflow is written before it runs: steps that act through tools, conditions for moving on, branches, retries on failure. This prior writing has consequences that are underestimated: the path followed is reproducible, therefore auditable and testable step by step, and a failure can be located. The cost becomes predictable too, since the number of calls to the model is known before anything is launched. You replay a workflow identically, and that is precisely what cannot be promised of an agent.

The decisive distinction

The decisive distinction from the agent bears neither on complexity, nor on the presence of a model, but on where the decision is taken. A step can perfectly well entrust a model with classifying a request, then route it to one of the planned branches: the decision stays inside the plan, written by a human. In an agent, the plan does not exist in advance, because the model observes the state of the work and chooses what comes next, turn after turn. Between the two lies a grey area, that of sequences whose structure is fixed but some of whose choices are delegated; it is that area which is most often called, a little hastily, agentic.

The traps

The first trap is paying for the unpredictability of an agent where a workflow was enough: if you know how to write the steps, writing them costs less and is easier to monitor. The second answers it in mirror image, because a workflow becomes unmanageable when exceptions multiply, each special case adding a branch until the plan is no longer understandable by anyone. It is this proliferation, and not the desire for autonomy, that honestly justifies moving to an agent. Finally, a share of the systems presented as agents are workflows: that is not a defect, but it changes what should be expected of them and what should be measured.