Fixed Workflows and Adaptive Agents

In the previous chapter, we focused on how the agent loop works. This chapter asks a prior question: Should we use an agent loop at all? The loop is not the default choice. It is one of three architectural options, and choosing the wrong one leads to systems that are expensive, fragile, or both.

In this lesson, we work through two concrete scenarios side by side. By the end, we will have a clear decision framework for choosing between a fixed workflow and an adaptive agent, and we will know when a task’s structure is the deciding signal. By the end of this lesson, we will be able to:

• Identify the structural difference between tasks that suit fixed workflows and tasks that suit adaptive agents.

• Apply the two-question decision test to any scenario.

• Name the primary reliability risk that each architecture introduces.

• Recognize the anti-pattern of using an adaptive agent when a fixed workflow is correct/

The architectural choice before the loop

Every Claude-powered system begins with one decision: how predictable is the task?

• If the steps are known in advance and they are the same steps in the same order every time, a fixed workflow (also called a prompt chain) is the right choice. Each step takes an input, produces an output, and passes it to the next step. No agent loop, no tool selection, and no runtime branching.

• If the path through the task depends on what is found at runtime, and step three can only be decided after seeing the result of step two, an adaptive agent is the right choice. The loop exists precisely to handle this: call the API, read the stop reason, and dispatch the next action based on what Claude decided.

The difference is not about capability. It is about task structure. Both architectures use Claude. Both can call tools. The choice depends on whether the next step is determined before execution (fixed) or during execution (adaptive). ...