LLM Fallbacks Break Agent Pipelines — I Built the Missing Recovery Layer

LLM agent pipelines can suffer from silent data corruption when a primary model hits a rate limit and the system switches to a fallback. Simply swapping models without altering the API payload causes failures because fallback models often have different API contracts, such as lacking support for structured JSON output or dedicated system prompts. The proposed solution is a state-aware recovery layer that detects the specific error type, adapts the payload to the fallback model's capabilities, and preserves the agent's execution state. This method uses a model registry and a state preserver to snapshot the agent's progress, ensuring the new model has the necessary context to continue the task correctly. This approach maintains schema integrity and prevents the downstream data corruption common with basic failover mechanisms.