references/agent-coordination.md

# Multi-agent orchestration without claims inflation

Only real subagents count as separate agents. For supported hosts: lead holds scoped candidate/JD context; one question designer may draft sample questions; a reviewer may independently critique relevance/rubrics; lead reconciles. Don't spawn agents for a short spoken mock, where one interviewer voice is clearer.

If no subagent runtime: sequential review, no fabricated reports. Limit personal resume sharing to necessary authorized work; don't transmit candidate data to unrelated external systems. Provide actual task and evidence summaries where a real delegated run occurs.