I have settled into a split that works for me. The agent collects evidence: reviews, competitor claims, search language. It sorts that into themes with sources. I write the argument and the recommendation.
When I let it write the argument too, every strategy came out as 'be more human and authentic'. Which is never wrong and never useful.
'Be more human' is the agent equivalent of 'make it pop'.
Same split, freelance version. The evidence pass used to take me two days. Now it takes an afternoon and I spend the saved day on the argument.
Same split for visual references. It finds and sorts, I choose. The choosing is the job.
The argument is the deliverable. The evidence is the appendix nobody reads but everybody wants to exist.
Research lead: please make the agent cite where each piece of evidence came from. Half the 'evidence' I am sent has no source.
Joined today, and this is the first thread I read. Design systems lead: we use the same split for documentation audits. The agent finds the inconsistencies, a person decides which ones matter.