The Do-I-Need-An-Agent Check

Help me decide whether this should be an agent at all.

The task: [DESCRIBE]
Current approach, if any: [DESCRIBE]
Volume: [HOW OFTEN, HOW MANY]
Cost of a wrong result: [DESCRIBE HONESTLY]
Who is affected when it goes wrong: [DESCRIBE]

Assess against four criteria:
1. COMPLEXITY — is this multi-step and hard to fully specify in
   advance? Or is it a well-defined transformation?
2. VALUE — does the outcome justify higher cost and latency than a
   single model call or plain code?
3. VIABILITY — is a language model actually good at this task type,
   or am I hoping it will be?
4. ERROR RECOVERY — can mistakes be caught and reversed? What
   catches them?

Then tell me plainly which of these is the best fit:
- Plain code, no model involved
- A single model call
- A fixed workflow with model calls at defined steps (my code
  controls the flow)
- A genuine agent (the model controls the flow)

Argue for the simplest option that could work. If an agent is not
justified, say so directly — do not design one anyway.

Then: if I build the simpler thing first, what would I learn that
would tell me whether the agent is worth building later?