Non-LLM Task Implementation

Not every workflow task needs an LLM — mechanical, retrieval, and classification tasks are usually cheaper and more dependable without one.

Before implementing any workflow task with an LLM, check whether it needs one at all:

Purely mechanical or retrieval tasks — crawling a site's content, saving content to a database — should just use traditional software. An LLM adds cost, latency, and nondeterminism for zero benefit over code that already does the job reliably.
Where classification suffices, use a dedicated classifier (e.g. a BERT-based model) rather than an LLM. A classifier is more dependable — it won't "make commentary" instead of classifying — and is faster and cheaper to run per item.

This is the task-level instance of the same principle progressive agent architecture applies at the system level: don't reach for the more general, more expensive tool when a narrower one already solves the problem. Deciding whether a monitored system has a problem at all is one such task: a deterministic score should gate any LLM diagnosis. Reserve templated-prompt and tool-based task implementations for the tasks that genuinely need an LLM's language understanding or generation.

Sources

Prompt Engineering for LLMs (John Berryman), ch. 9