AWS builds NarrateAI, a business analytics assistant on Amazon Bedrock AgentCore

AWS has unveiled the architecture of NarrateAI, a conversational assistant built on Amazon Bedrock AgentCore that helps the SMGS (Sales, Marketing and Global Services) division scale its business analytics. The system combines batch processing with real-time interaction to deliver insights across the organization.

Two-layer architecture

NarrateAI splits processing into two independent layers. The first layer runs batch jobs: data preparation, metric calculation and result caching. This delivers high throughput and saves compute resources. The batch layer works asynchronously and parallelizes the processing of large datasets.

The second layer handles users' real-time requests: conversational interaction, redirection to the right agents and input validation. It is responsible for latency and answer quality at the moment of the request.

Amazon Bedrock AgentCore makes it easier to manage orchestration between the layers and the parallel execution of specialized agents. This split is critical for scaling: the batch layer copes with growing data volumes, while the real-time layer keeps latency under control and is not blocked by long-running computations.

Specialized agents

Instead of a monolithic assistant, AWS uses a network of specialized agents, each with its own area of responsibility. The validation agent checks that an incoming request is correct and complete before it enters the main pipeline. The routing agent decides which SMGS unit (sales, marketing, global services) should receive the request, taking context into account. Knowledge agents supply context specific to each unit: metrics, historical trends and internal processes. The synthesis agent merges answers from different sources into a coherent recommendation that the user sees.

This architecture lowers the error rate, improves answer relevance, makes it simpler to add new agents and eases debugging.

Engineering patterns for production

AWS highlights several practices that allowed it to launch NarrateAI at scale:

• Graceful degradation: if an agent is unavailable, the system offers an alternative answer instead of an error
• Cost tracking: each agent tracks its processing cost, and the system optimizes routes
• Observability across the agent chain: logging, tracing and latency monitoring at every node
• Rate limiting and prioritization: the batch layer does not block real-time requests, and queues are separate
• Caching of agent responses for frequent requests

These patterns helped AWS achieve high reliability and predictable behavior in production.

What this means

The article shows that production-quality AI systems need more than an LLM: they need an architecture that handles scale, latency and reliability. Specialized agents, two-layer processing and graceful degradation are practices that other companies can apply with Amazon Bedrock. For developers, it means AI orchestration is becoming a separate layer between the LLM and business logic.