MoEngage used AWS’s AgentCore Gateway to make roughly 400 private APIs available as tools for AI agents, while keeping them off the public internet and controlling which tools each agent can access. The case study offers a practical look at how a company can connect agents to existing services without handing a model the keys to the whole workshop.
AWS AI Watch analysis
What happened
The APIs serve MoEngage’s customer-engagement platform and are maintained by 25 engineering teams across seven AWS regions. In a post co-written with MoEngage engineers, AWS says the company used AgentCore Gateway to turn API specifications into agent tools, with Cedar policies filtering each agent’s available tools on the server side.
MoEngage also keeps tenant and user identity out of the model’s tool arguments, passing it through request headers instead. AWS says the approach saved an estimated six months of platform-engineering effort, and that onboarding a new API now takes hours rather than days. Those are estimates reported in the case study, not independently measured comparisons.
Why it matters
The hard part of putting agents to work inside a business is often not inventing new functions. It is exposing existing ones without opening a security, networking or context-window headache. Here, agents can discover relevant tools without receiving a catalogue of hundreds, and access rules are enforced before tools are shown or called.
That is a useful implementation pattern for teams weighing a central gateway against building and maintaining a separate MCP server for every service. It also shows where the control boundary sits: identity and permissions are handled outside the model, rather than entrusted to a prompt with excellent manners.
Our read
This is a concrete, technically useful example of AgentCore being applied to a real API estate, not just a diagram in search of a deployment. The most persuasive details are the server-side tool filtering and the decision to keep identity beyond the model’s reach. The time savings are promising, but remain MoEngage’s estimate; the operational trade-offs are worth checking before copying the architecture wholesale.
What to watch
- Whether MoEngage shares measured results on latency, cost or operational workload.
- How its Cedar policies are maintained as APIs and agent permissions change.
- Whether the pattern proves workable for teams with different networking and identity requirements.
Discussion spark: Would you trust a centrally managed gateway to set agent permissions across dozens of engineering teams, or should each team control its own tools?
Sources and evidence
- How MoEngage turned over 400 APIs into agent tools with AgentCore Gateway (29 October 2026, 16:54 UTC)
not affiliated with or endorsed by Amazon Web Services (AWS)