Build the research and enrichment agent against your one-pager: a list of real targets in, verified structured fields out, every claim linked to its source, under the budget you set. Design the schema before the prompt, and the walls before the model: plain code decides what never needs a model call, and validates everything the model returns.
Friday is your first live debrief: the cohort's predictions against the cohort's results, and the first failure gallery.
It ends with
Verified accuracy and cost figures in your repo, checked against a hand-labeled sample you built yourself.