An agent or a fixed pipeline

Should the model decide what to do next in a loop, or should you write the steps and call the model at fixed points?

The short answer

Write the pipeline. Reach for a loop only where you genuinely cannot name the next step — and then bound its steps, its tokens and its wall clock.

Fixed pipeline

Good at

Predictable cost and latency, because the number of model calls is known before you start.
Debugging — a failure is at a named step, not somewhere in an eleven-turn trace.
Testing each step independently, with ordinary unit tests.
Meeting a latency budget, since you can parallelise steps you know are independent.
Safety: a step that cannot be reached cannot be misused.

Costs you

Every unanticipated input shape needs new code, not a better prompt.
Branching logic grows into a decision tree nobody wants to maintain.
Cannot recover from its own mistakes — a bad extraction flows through to the end.
Worse on genuinely open tasks where the right sequence depends on what is found.

Pick this when: The task has a shape you can draw. Most production LLM features are this, including several that were built as agents first.

Good at

Tasks where the next step depends on what the last one returned.
Recovering from failure: a tool error is an observation, not an exception.
Long-tail inputs a fixed flow would need dozens of branches for.
Search-shaped work — debugging, research, codebase changes — where the path is the problem.

Costs you

Cost and latency are unbounded until you bound them, and a loop can burn a lot of tokens.
Errors compound: 95% per step is about 60% over ten steps.
Evaluation is much harder — you have to grade a trajectory, not an answer.
A wider attack surface, because the model chooses the tool calls.
Non-determinism makes reproducing a bug report genuinely difficult.

Pick this when: You cannot enumerate the steps in advance, and the task tolerates variable cost, latency and outcome.