Automate the boring path before you reach for agents
Most work described as needing an agent is a deterministic workflow with three exceptions. Build that first.
Agents are genuinely useful for work that can’t be drawn as a flowchart. The problem is that most of the work people want to hand to an agent can be drawn as a flowchart, and would be cheaper, faster and far more reliable as one.
A simple test
Write the process down as steps. If you can complete it without the phrase "it depends on what comes in", it’s a workflow. Build the workflow.
Where language models still earn their place in a workflow
- Reading unstructured input and turning it into structured fields
- Classifying an inbound message into one of a known set of categories
- Drafting a response for a human to approve
- Summarising a long document against a fixed set of questions
In each of those, the model does one bounded job inside a deterministic system. The system decides what happens next. That’s the difference between something you can run on live operations and something you have to supervise.
The reliability of a system is the reliability of its least deterministic step, multiplied across every step.
Start deterministic. Add judgement only where the flowchart genuinely breaks down, and put a human in front of anything irreversible.