Sam Prodger compares two hotel-booking agents: one relies on instructions and exposed tools, while the other uses a scoped identity and gateway policies. In his demonstration, the governed agent can cancel an authorized booking while a bulk-deletion request is denied. Another example blocks access to mocked guest data rather than accepting a user's compliance explanation as authorization.
Sam Prodger then demonstrates cost controls before inference. A rate limit admits three requests per minute and stops a repeat loop; semantic caching serves repeated booking questions without another model call. An oversized prompt is rejected before downstream processing. These are specific vendor demonstrations, not evidence that a gateway alone eliminates every agent-security risk.
Sam Prodger also discusses model routing, smaller MCP tool catalogs, exportable telemetry and visibility into AI traffic on employee laptops. The reusable lesson is to put enforceable permissions and resource limits in the infrastructure that executes requests, while retaining logs that show what was allowed or blocked. Booth invitations and broader promotional claims are omitted.
Watch on YouTube
