Most golden dataset guides assume one team scoring one flat input-output pair. An AI agent needs the full trajectory, tagged per tenant, and tied to the number your own customers see.
Shadow testing runs a candidate AI agent version against real traffic before it ships, but comparing two non-deterministic outputs and stopping duplicate tool calls take real engineering the generic guides skip.
Prompt versioning tracks changes to an AI agent's prompt, but an eval score that passes on average can still hide a regression that only shows up in one tenant's deflection rate.
AI agent hallucination is not always a wrong fact. Often it's a confident claim that no tool call ever backed up, and it skews the deflection rate you report to customers.
Prompt injection testing for AI agents has to check tool calls and tenant boundaries, not just chat replies. A concrete test method with real examples.
AI agent testing catches tool failures and bad outputs before they land in the deflection and cost numbers you report to customers. Here is what to test first, with real failure examples.