When should you use a fixed DAG pipeline instead of a dynamic agent loop?

The choice between a **fixed pipeline** and an agent comes down to one question: who decides what runs next? In a fixed pipeline, also called a chain or a DAG, you the developer decide the steps in advance. Step A always feeds step B, which always feeds step C. The flowchart is drawn before the task starts and never changes. This is predictable, cheap, and easy to debug, because the same path runs every single time. In an agent, the model decides what runs next. After each step it looks at the result and chooses the next action on the fly. That adaptability is powerful, but it costs an extra model call on every turn and makes latency and cost hard to predict. So you pick the fixed pipeline when the execution path is fully known at design time. If you can draw the whole flowchart up front with no conditional branches, you do not need an agent. You reach for an agent only when the path genuinely depends on what intermediate steps return, and you cannot know that ahead of time. The wrong answers in the question describe complexity, model weakness, or flaky APIs. None of those is the real test. The real test is fixed versus dynamic control flow.

Use a **fixed DAG pipeline** when the execution path is fully determined at design time and you need predictable latency and cost. Use an agent loop only when the next step must depend on intermediate results you cannot know in advance. The defining distinction is control flow. In a chain, the developer hard codes the sequence of steps. In an agent, the model decides each next step at run time. That single difference drives every tradeoff. A fixed pipeline wins on three properties. Latency is bounded because the step count is fixed. Cost is bounded for the same reason. Debugging is simple because the same path runs every time. These are exactly the guarantees production systems want. An agent earns its keep only when the path genuinely branches on what a step returns. Multi-hop research, where each finding changes the next query, is a fit. So is open ended coding, where the next edit depends on the last test result. The trap is choosing an agent because a task feels complex. Complexity is not the test. A long pipeline with twenty deterministic steps is still a chain. The test is whether you can draw the full flowchart before the task runs.