From first-time success to stable operation
Prototypes usually verify core capabilities, and production systems also have to deal with concurrency, timeouts, model unavailability, interface exceptions, and data changes.
Bring identity and permissions into every call
Models, knowledge retrieval, and tool invocation all need to know the current user identity and follow the permission boundaries of the existing business system.
Create repeatable quality measures
Measurement sets should be constructed using real business problems and continuously checked for accuracy, referencing, formatting, rule compliance, and task completion.
Make the running process observable
Record request time, model versions, knowledge sources, tool calls and user feedback to locate problems and measure improvements.
Preserve manual intervention and fallback paths
For uncertain, high-risk or abnormal situations, personnel should be allowed to take over the task and be able to return to the original business process.