---
type: "post"
id: "how-moengage-built-agentic-customer-engagement-on-amazon-5c08ee"
slug: "how-moengage-built-agentic-customer-engagement-on-amazon-5c08ee"
title: "How MoEngage built agentic customer engagement on Amazon Bedrock AgentCore"
description: "AWS’s MoEngage case describes Merlin AI, a set of agents that lets marketing and product teams work with campaigns through conversation."
retrieval_nugget: "AWS’s MoEngage case describes Merlin AI, a set of agents that lets marketing and product teams work with campaigns through conversation. The notable design choice is to expose existing OpenAPI services as MCP tools while keeping identity, tenant context, and session isolation outside the model—an implementation pattern for software teams making established products agent-accessible."
published_at: "2026-11-17"
updated_at: "2026-11-17"
record_date: "2026-11-17"
date_kind: "published_at"
topics: ["claude","anthropic"]
entities: ["MoEngage","Merlin AI","Amazon Bedrock","Amazon Bedrock AgentCore","FastMCP","Anthropic","Claude Sonnet 4.6","Glance"]
related_patterns: ["agent-ready-software-exposes-capabilities"]
novelty_label: "new_fact"
what_changed: "MoEngage’s customer-engagement case study adds a distinct application of AgentCore, involving Merlin AI and customer engagement rather than exposing APIs as agent tools."
prior_refs: []
source_url: "https://aws.amazon.com/blogs/machine-learning/how-moengage-built-agentic-customer-engagement-on-amazon-bedrock-agentcore"
source_urls: ["https://aws.amazon.com/blogs/machine-learning/how-moengage-built-agentic-customer-engagement-on-amazon-bedrock-agentcore"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:how-moengage-built-agentic-customer-engagement-on-amazon-5c08ee"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/how-moengage-built-agentic-customer-engagement-on-amazon-5c08ee/","markdown":"https://newruntime.com/posts/how-moengage-built-agentic-customer-engagement-on-amazon-5c08ee.md","json":"https://newruntime.com/posts/how-moengage-built-agentic-customer-engagement-on-amazon-5c08ee.json"}
markdown_url: "https://newruntime.com/posts/how-moengage-built-agentic-customer-engagement-on-amazon-5c08ee.md"
---

# How MoEngage built agentic customer engagement on Amazon Bedrock AgentCore

## Retrieval answer

AWS’s MoEngage case describes Merlin AI, a set of agents that lets marketing and product teams work with campaigns through conversation. The notable design choice is to expose existing OpenAPI services as MCP tools while keeping identity, tenant context, and session isolation outside the model—an implementation pattern for software teams making established products agent-accessible.

## Summary
AWS’s MoEngage case describes Merlin AI, a set of agents that lets marketing and product teams work with campaigns through conversation. The notable design choice is to expose existing OpenAPI services as MCP tools while keeping identity, tenant context, and session isolation outside the model—an implementation pattern for software teams making established products agent-accessible.

## What happened
AWS’s announcement, co-written with MoEngage employees, says MoEngage built Merlin AI on Amazon Bedrock and Amazon Bedrock AgentCore. The platform is used by more than 1,350 consumer brands across channels including push, email, in-app, SMS, WhatsApp, and Rich Communication Services. Merlin AI agents have been in production for more than six months, according to the post.

The system has four agent modes. Embedded agents in Merlin AI Studio support tasks such as writing personalized copy, creating banners, translating audience descriptions into segment filters, and producing Jinja or Liquid markup. Conversational agents answer questions and act across campaigns, analytics, segments, and flows; the Campaign Insights Agent can delegate analysis, diagnosis, and comparisons to specialized sub-agents. Custom agents run continuously on MoEngage data and tools within rules chosen by teams, with an activity log and a choice between autonomous execution and human review. Decisioning agents select content, timing, frequency, and channel for individual customers, with goals and guardrails set by product and marketing teams.

Rather than rebuild its APIs, MoEngage used existing OpenAPI 3.0-compatible specifications with FastMCP to generate MCP tools. The tools cover campaign operations, analytics, segmentation, content, and flows. Changes to the specifications keep the tool definitions aligned with the APIs. The MCP server runs on AgentCore runtime and calls the same platform APIs customers already use; it does not run its own model. Amazon Bedrock models handle reasoning and generation. The post names Anthropic Claude Sonnet 4.6, Claude Sonnet 5, and Claude Sonnet 5.5.

For a typical request, a signed-in user’s existing JSON Web Token is validated by AgentCore runtime, which starts an isolated session. Workspace and user context travel in request headers, not as tool arguments visible to the model. The server calls MoEngage APIs with the user’s tenant context, and those APIs enforce existing permissions and data-residency rules. AWS Bedrock Guardrails filter content and personally identifiable information in responses; a backend-for-frontend layer validates generated content before the interface renders it. Users can review actions and require confirmation before execution—for example, the campaign flow described in the post pauses an SMS branch only after confirmation.

## What the archive adds
New Runtime’s earlier coverage on [building a semantic layer for agentic AI on AWS](https://newruntime.com/posts/build-a-semantic-layer-for-agentic-ai-on-ec1cc0/) (October 22, 2026) argued that governed definitions can provide a stable boundary between enterprise data and agents. This is a new customer implementation, not a new fact about that earlier project: MoEngage’s case applies the boundary idea through existing APIs, tenant context, and MCP tools.

## Trend signal
This case supports [agent-ready software exposes capabilities](https://newruntime.com/patterns/agent-ready-software-exposes-capabilities/): MoEngage makes platform operations composable through structured tools generated from existing API specifications. Its access design also illustrates why exposed capabilities need explicit permissions and isolation, rather than relying on a model to respect tenant boundaries.

## Why now
The post describes a system already in production for more than six months, and presents the work as a response to marketing teams navigating analytics, content, and campaign tools separately. The timing is therefore a public account of an operating architecture, rather than a newly announced product launch.

## Why it matters
For a software team with a mature API surface, this case offers a route to agent access without duplicating business logic in a new agent-specific backend. Generating tools from OpenAPI specifications can keep the agent interface connected to the same operations and validation already used by customers. More importantly, the design assigns trust decisions to established system boundaries: authenticate the user before tool execution, keep tenant context outside model-editable arguments, preserve API permissions, and isolate sessions.

The case also distinguishes between different levels of autonomy. A team can allow an agent to work continuously under configured rules, inspect its activity, or require human review before consequential actions. That is a more useful operational choice than treating “agentic” as synonymous with unrestricted execution. The reported campaign and conversion improvements may be encouraging, but teams should treat them as hypotheses to test against their own baselines and workflows.

## Limits and uncertainty
AWS and MoEngage report the results, but the post does not provide the underlying measurement methods, comparison groups, sample sizes, or customer-level breakdowns for the platform-wide figures. Glance’s reported 38% push click-through-rate uplift and 50% faster campaign go-live are not independently verified here. The post does not quantify latency, error rates, tool-call failure handling, or the human-review burden. AgentCore feature availability varies by AWS Region, and the account does not give a complete deployment or pricing breakdown for reproducing the architecture.

## Skill to build now
Take one stable operation from an existing API and test whether its OpenAPI specification can generate a useful MCP tool. Keep authorization claims out of model-editable tool arguments: validate the caller’s existing identity, pass tenant context through trusted server-side context, and have the underlying API enforce its existing permissions. Before letting the tool execute changes, add an explicit review step and an activity log so a person can inspect what the agent read and attempted.

## Reader takeaway
MoEngage’s reusable lesson is architectural: expose capabilities from existing APIs, but do not make the model the authority on who can use them or which tenant’s data it sees. Tool generation, identity validation, API permissions, isolated sessions, and a deliberate human-review option work together to make conversational access operationally safer.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Claude](https://newruntime.com/topics/claude.md): Explore the claude topic hub.
- topic hub: [Anthropic](https://newruntime.com/topics/anthropic.md): Explore the anthropic topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)
