Workflows vs Harnesses
Agentic systems run along a spectrum. At one end are fixed workflows; at the other, open agent harnesses. What changes along it is one thing: how much the model decides for itself.
Fixed Path or Open Loop
The same tools sit at both ends. The difference is who chooses the next step.
- Workflow — a predefined sequence of LLM calls that you design and wire together.
- Agent harness — a loop where the LLM repeatedly decides its own next step until the task is done.
Control vs Flexibility
Neither end is better — each buys something at a cost.
Let the Task Decide
The task tells you which end to reach for.
- Known steps — if you can list the steps up front, use a workflow.
- Unknown or branching path — if the steps depend on what the model finds, use a harness.
- Most real systems mix both — workflows with a small harness where the path is genuinely open.
Start as Fixed as You Can
Add autonomy only where the task truly needs the model to decide. Every step you hand to the model is a step you can no longer predict — so more autonomy means more to guard against, a theme Lesson 6 returns to.
Build It
How to implement: for one feature of your build, decide — is the path fixed (a workflow) or does the model need to choose its own steps (a harness)? Write down which, and why.
- Weekly AI Tasks tracker — parsing a message into a task is a fixed workflow (extract → validate → store); answering “what should I focus on?” leans toward a small harness.
- Personal brand site — generating each section is a workflow; no open-ended agent needed.
What you learned
A workflow is a sequence you design; a harness is a loop the model drives. Workflows trade flexibility for control; harnesses do the reverse. Pick by the task, mix where it helps, and start as fixed as you can.