Interesting implementation. We have had better reliability by treating tool access as a capability-routing layer: request by task, not fixed provider, with explicit cost/latency constraints and runtime trust state (verified, reachable, failed). It keeps orchestration logic out of agent code and makes fallback behavior auditable.