Developers building MCP agents, when processing candidate content (items to ingest, tickets to answer, results to rank), must first verify, screen, classify, rerank, or score it, then decide whether it is auto-approved, sent to human review, or escalated.
Public materials do not state the current practice; by workflow inference, developers today write their own prompts and validation code, or hard-code pass rules in the business system.
The public description splits judgment from release policy, implying the pain: when agents let the model judge directly, release criteria are scattered in prompts, thresholds cannot be tuned to business risk, and errors cannot be traced to a specific judgment step.