The Workflow vs Agent Decomposer

Help me split this problem into the parts that need an agent and
the parts that do not.

The full problem: [DESCRIBE]
Current design: [DESCRIBE, OR "none yet"]

Break it into steps. For each step, classify:
- DETERMINISTIC: plain code, no model
- SINGLE CALL: one model call with a fixed prompt
- CONSTRAINED: model call with structured output, validated
- AGENTIC: model decides the next action from the result of the
  last one

Then produce:
1. The decomposition, with each step classified and justified
2. Which steps I have wrongly classified as agentic that are
   actually deterministic
3. The control flow: what my code orchestrates vs what the model
   decides
4. Where the boundaries between deterministic and agentic parts are
   — these are where errors accumulate
5. What each boundary must validate
6. A simpler version of this design that would still deliver most
   of the value

The default should be: as little agentic control flow as the
problem genuinely requires.