For any multi-agent design discussion in the RRK, you should touch:

Reliability

• Timeout per agent call (don't let one stuck agent hang the whole system)
• Retry with exponential backoff on transient errors; circuit breaker on persistent
• Max-step cap on agent loops to prevent runaway

Latency

• Parallelize independent agent calls (fan-out/gather)
• Stream supervisor output even while workers are still running where possible
• TTFT optimization: smaller model for the streaming "thinking out loud" layer

Cost

• Model routing per role (Flash for triage, Pro for synthesis)
• Aggressive context pruning between agents — don't pass the whole history downstream
• Cache tool results (semantic cache for retrieval, exact cache for deterministic tools)
• Per-agent token budgets + circuit breakers

Observability — the hardest one for multi-agent

• Distributed tracing across agent boundaries (one trace ID per user request, spans per agent + per tool call)
• Per-agent token + cost attribution (which agent is burning budget?)
• Outcome metrics (did the workflow complete? was the final answer correct?) separately from process metrics (which agents fired, in what order, how long each took)

Evaluation

• End-to-end task success rate is the north star