Multi-provider LLM router with predictive rate-limit avoidance for production agent runtimes

“Datadog's Feb 2026 telemetry showed 5% of LLM call spans erroring with 60% from rate limits; March 2026 dropped to 2% but rate limits stayed near a third (https://www.datadoghq.com/state-of-ai-engineering/, https://www.braintrust.dev/articles/agent-observability-complete-guide-2…”