AWS published a case study on 29 October describing how MoEngage, a customer-engagement platform used by more than 1,350 brands, made roughly 400 internal REST APIs reachable by its AI agents. The APIs are built by 25 engineering teams and run in private VPCs across seven AWS Regions for data-residency reasons, with no public internet exposure. MoEngage rejected the obvious MCP route - one server per team - because it would mean more than 20 deployments to patch and keep protocol-compatible, and cross-domain agents would have to reconcile several servers. Instead the AI platform team owns a single AgentCore Gateway: product teams describe operations in an OpenAPI spec, the gateway exposes selected operations as tools, and per-agent authorisation is enforced server-side through Cedar policies tied to each agent's IAM role. Tools are discovered by describing the task rather than by loading every definition into context. The setup has run in production for three months, serving about 400 registered operations to ten internal agents. MoEngage estimates it saved roughly six months of platform work and that onboarding a new API now takes hours of spec annotation instead of days.
Our takeThe transferable idea is search-first tool discovery: an agent describes what it wants, gets a shortlist, and the context window holds candidates instead of hundreds of definitions, while renamed operations stop breaking agent code. That moves the quality burden onto the OpenAPI descriptions - vague operation summaries will now degrade tool selection directly, and there is no separate routing layer to compensate. Note also what is and is not demonstrated: private connectivity and Cedar-enforced scoping are infrastructure properties, while the six-month saving is the company's own estimate rath