A fire department wants a runbook assistant where research work stays separate from dispatch suggestions. Which multi-agent choice supports that?

The fire department runbook assistant where research work stays separate from dispatch suggestions should use separate research and dispatch-suggestion agents with clear role separation. Research agents gather grounded runbook context; dispatch-suggestion agents propose options without owning unconstrained computer-aided dispatch writes. Role separation under multi-agent design reduces confused responsibilities during high-pressure incidents and keeps research quality distinct from operational recommendations. The wrong choice of separate research and dispatch-suggestion agents with distinct model deployments misses the required separation because distinct deployments only change hosting, which doesn't keep research work separate from dispatch suggestions in role or behavior. The wrong choice of separate research and dispatch-suggestion agents with dedicated token usage quotas misses the required separation because token quotas only limit usage cost, which doesn't keep research reasoning distinct from dispatch recommendations. The wrong choice of separate research and dispatch-suggestion agents with staggered retry intervals misses the required separation because staggered retries only change timing on failure, which doesn't keep research work separate from dispatch suggestions. On the exam, separate research vs dispatch-suggestion agents is the multi-agent choice; wrong choices are distinct model deployments, dedicated token usage quotas, or staggered retry intervals. Caveat: separation is not isolation—define a handoff so dispatch suggestions consume research outputs without letting the research agent write to computer-aided dispatch. Operational check: run a runbook query through the research agent into the dispatch-suggestion agent, verify computer-aided dispatch write tools are absent from both or tightly gated, and confirm suggestions cite research context.