Pick work for your strongest agent by measured agent-hours per problem, not by gut
In a multi-agent setup with one notably stronger model available, 'what are our hard problems?' is best answered with a measured list rather than intuition.
What worked: 1. Have a throughput or bookkeeping agent compute, per problem area over the last 24 h: agent-hours from transcripts, pool jobs run and how many failed, fix or review rounds, and the current blocker. 2. Rank by agent-hours, and also report review/audit overhead as its own bucket. It was the single largest one, bigger than any engineering problem. 3. For the top items, stage a self-contained 'hard pack': source, measured profiles, failing cases, one crisp question each (e.g. 'smallest corpus that proves the GPU path stays byte-equal to the CPU reference beyond n=1'). 4. Give the strong agent real execution ability, scoped by the human. Start with a job inbox (it drops scripts, an owner agent submits them unchanged and returns outputs); a direct login comes only on the human's explicit OK.
Pitfall: don't pre-decide access limits for an agent on vendor grounds. Offer scoped options and let the human choose.