A research agent is built with a fixed pipeline: Step 1, search. Step 2, read top 5 results. Step 3, summarize. A teammate proposes converting it to a ReAct pattern — the agent reasons about what to search for, observes what came back, and decides what to do next based on what it found.
The team is debating: when does ReAct beat the fixed pipeline?
What's the right framing?
Why did you pick that answer? Two or three sentences. The act of articulating it is what builds the judgment — not the click that follows.
ReAct fits when the workflow needs to adapt to discovered information; fixed pipelines fit when transformations are predictable. A finance pipeline (categorize → validate → submit) doesn't need ReAct — every step's input is known. A research workflow (search → assess → search again with different terms if needed → assess → synthesize) absolutely does — what to do at step N depends on what step N-1 returned. The decision rule isn't "ReAct or pipeline"; it's "is this workflow predictable or adaptive?"
Dynamic decomposition has costs: unpredictability, harder debugging, higher latency, more failure modes. For predictable workflows where the steps are known, fixed pipelines are more reliable, not less. "Always more flexible" sounds like a virtue but is actually a non-criterion.
"Predictable equals reliable" is true for predictable workflows and false for adaptive ones. A research pipeline that rigidly reads 5 papers when only 2 exist is predictably doing the wrong thing — predictability without fitness for purpose is brittleness.
This is the genuinely tempting wrong answer because it sounds disciplined — "if our prompt anticipates the cases, we don't need adaptive control flow." The problem is combinatorial: every edge case you anticipate adds prompt length and complexity, and the cases you didn't anticipate still surprise you. ReAct is the architectural answer to open-ended situations where listing every case isn't tractable. "Better prompt" is appropriate when the cases are bounded; ReAct is appropriate when they aren't. Treating them as substitutes is a category error.