Part 2 of 5 · 2 min read
Step zero and the checklist
Check whether you need a model at all, then answer six questions and see which way they lean.
Before comparing, check the cheaper options: a rule, a regular expression, a database query, a lookup table. If inputs are predictable and structured, ordinary code is faster, cheaper and more reliable. A model earns its place where input is unstructured text, images or audio, or where the rules are too fuzzy to write.
recommend() turns six yes or no questions into a leaning, and prints its reasons so you can argue with it:
| Question | A yes pushes toward |
|---|---|
| Can you write down the steps for almost every input? | workflow (+2) |
| Do inputs fall into a small set of kinds? | workflow (+2) |
| Must the next step depend on tool results in ways you cannot list? | agent (-3) |
| Is a wrong action expensive or hard to undo? | workflow (+1, more control) |
| Is there a strict latency or cost limit? | workflow (+2) |
| Must you explain exactly why each step happened? | workflow (+1) |
recommend(dict(steps_known=True, paths_enumerable=True, adapts_to_results=False,
mistakes_costly=True, tight_budget=True, needs_audit=True))
# Verdict(leaning='workflow', score=8, ...)
The score decides the leaning: 2 or more means workflow, minus 1 or less means agent, and 0 or 1 means "start with a workflow, add one agent step where needed". The weights are my judgement, not a law. The value of the function is that it forces you to answer each question and shows which answer decided it. Misspelling a question name is reported as an error rather than ignored.
Note the mixed case: "we know the steps" (+2) with "the next step depends on tool results" (-3) scores minus 1, which leans agent. That is deliberate. If the next step genuinely depends on what a tool returned in ways you cannot list, a fixed path will fail on those cases, however well you know the usual ones. The hybrid is then the safer way to build it.
Check yourself
A task has known steps, a tight budget, and mistakes that are costly, but one stage needs to adapt to search results. What do you build? (A workflow, with that one stage as a capped agent. The three workflow answers score +5, and the adaptation answer adds -3.)