An AI gateway is an intermediary layer for governance and operational controls between AI systems and services. In agent workflows, it can maintain running counts across a session so that controls can account for more than one action. section 3.8
Runtime rules can be written as policy code and checked before each action. Counts add context for rules that depend on repeated behavior, rather than on a single tool call. section 3.8
Monitoring a clinical agent
A clinical agent may make one emergency escalation per patient per episode without review; a second requires a person. If it makes more than three times the tool calls of a normal case, the product owner is notified. Such rules require the system to track activity over time. section 3.8
Sequence and delegation need additional safeguards: logs should identify both the executing agent and the requester chain, while authority must narrow at each delegation step. OAuth token exchange can carry caller context, but does not enforce permission narrowing by itself. section 3.8
