The Enterprise GenAI Reliability Checklist.

Your model provider will have outages and your RAG store will time out, so plan for both.

• 2.1 Every LLM call has a timeout, and you know what each one is set to.
• 2.2 Every agent loop has a maximum iteration count that prevents infinite tool-calling loops.
• 2.3 At least one fallback model is configured for critical user flows (e.g., GPT-4o primary, Claude secondary). Failover is tested under load.
• 2.4 Tool failures are catchable at the agent level. The agent degrades gracefully ("I couldn't complete the web search, but here's what I know").
• 2.5 Retry logic has bounded backoff and honors retry-after headers from providers, so you don't end up DDOSing Anthropic at 3am.
• 2.6 Long-running operations (document processing, embedding generation) run asynchronously via a job queue, outside the request path.
• 2.7 There is a tested runbook for "our primary LLM provider is down," and you have run it as a drill.