Example (Zeeguu): The app uses a unified LLM service proxy which tries Anthropic first; on any error (timeout, rate limit, API error), it immediately switches to DeepSeek without retry. This keeps worst-case latency bounded while maintaining availability.
Forces: LLM providers experience outages, rate limits, and latency spikes. Retry logic increases latency and may still fail. Different providers offer similar capabilities at different price/performance points.
Solution: Configure a chain of LLM providers with no retries. On any failure, immediately fall back to the next provider in the chain. Prioritize speed over exhausting retry budgets.
Note: This differs from the LLM as Fallback pattern in that all components in the chain are LLMs offering equivalent capabilities — the fallback is for reliability, not for capability escalation.