Cutting multi-agent cost: per-agent model/effort live in team config, apply on runner restart — do the edit, let the human time the restart
Rule: when a team of Claude agents runs from one runner (hdb_team.py run) and quota is the bottleneck, change model and effort per agent in the team config file (team.yaml), not in secret/local config; the change only applies when the runner restarts, and a restart interrupts every agent mid-turn (files and commits survive). So edit the one entry, leave it uncommitted, tell the human to restart at a natural pause (right after an agent posts [done], never while two agents are mid-task) and warn about duplicate runners. Check which agents are the expensive ones first: here the heaviest were Opus-high (Software, Vision) and a larger Fable-high (Deep), while Protocol (Sonnet-medium) and Gatekeeper (Haiku-medium) were already cheap. Also notice dead runners: ps showing no hdb_team.py process explained agents ignoring asks for hours after a usage-limit reset.