Amazon Web Services (AWS) has unveiled a serverless gateway designed to streamline how artificial intelligence (AI) agents access vast enterprise data lakes. This new architecture, deployed on Amazon Elastic Container Service (Amazon ECS) with AWS Fargate, aims to resolve the persistent challenge of integrating multiple AI agents with diverse datasets without duplicating code or compromising security.
The Chenab Times has learned that the system acts as a centralized, agent-agnostic bridge between AI agents and an organization’s data lake, simplifying the process of exposing new datasets. Traditionally, AI teams have directly connected agents to databases and APIs, leading to complex integration logic and fragmented access control that becomes unwieldy as the number of datasets or agents grows.
This novel approach, piloted by Amazon’s North America Sort Center (NASC) network planning teams, allows conversational access to operational data. By employing a metadata-driven strategy, the system significantly reduces the engineering effort required to onboard new data sources, transforming a multi-day task into a process requiring only a simple JSON metadata file.
Addressing Fragmentation and Onboarding Challenges
The development addresses two primary challenges: the fragmentation of data access patterns and an unsustainable onboarding model for new data. Enterprises often possess thousands of datasets across operational metrics, business analytics, and domain-specific information. When multiple teams seek AI assistants capable of querying this data conversationally, each integration is frequently built from scratch, embedding custom SQL generation logic, schema knowledge, and connection handling.
This custom approach leads to significant overhead. A schema change can necessitate patching every individual agent, and fragmented access control lacks a single audit point for data access decisions. Onboarding a new dataset typically requires custom tool code, query validation, and redeployments, often consuming several days per table. To counter this, the design established four key requirements: configuration-driven data lake access, agent-agnostic tool interfaces, centralized permission enforcement, and elastic infrastructure to support fluctuating AI workloads.
Solution Architecture: A Layered Approach
The solution separates concerns into four independently scalable and deployable layers: client, MCP server, tools, and knowledge-base refresh. A metadata catalog underpins the tools layer, storing schema descriptions, column-level semantics, and table relationships. This allows AI agents to derive schema knowledge at query time, eliminating the need for hardcoded table structures or redeployments during schema changes. The knowledge-base refresh layer ensures this catalog remains current by indexing new dataset metadata as it becomes available.
A request originates from a client application, such as a web chat or an IDE plugin. It routes through Amazon Route 53 and an Application Load Balancer, protected by AWS WAF, before reaching Amazon Cognito for authentication. Authenticated traffic then enters a private subnet where the client layer runs on Amazon ECS Fargate. This layer manages session state in Amazon DynamoDB and forwards requests to the MCP server layer, also on Amazon ECS Fargate, via an internal Application Load Balancer.
Within the MCP server layer, an AI agent, powered by Amazon Bedrock, orchestrates tools built with FastMCP, a Python framework for developing MCP servers. Amazon Bedrock Guardrails are applied to all model interactions, ensuring content safety and security. The data layer comprises tool backends like Amazon Athena and Amazon Redshift for SQL queries against the data lake, and Amazon Bedrock Knowledge Bases for semantic search. Data ingestion into Amazon S3 and knowledge bases is handled by AWS Lambda.
Metadata-Driven Access and Security
The MCP client layer handles authentication and session management using Amazon Cognito and DynamoDB. A context mapping configuration scopes each query to relevant tables, contributing to interaction latencies in the 15–20 second range during testing. Adding a new business domain simply requires updating this mapping configuration.
The MCP server layer, isolated on its own ECS Fargate cluster, manages query analysis, tool selection, execution, and response generation. It exposes a JSON-RPC interface via the MCP protocol endpoint (/mcp), allowing any MCP-compatible client to discover and invoke tools directly. Fargate tasks are stateless, with conversation state persisted in DynamoDB, ensuring data integrity during scaling or task replacement.
The tools layer comprises registered tools discoverable at runtime. These include SQL query tools utilizing Amazon Athena and Amazon Redshift for data lake access, and document and metadata search leveraging Amazon Bedrock Knowledge Bases for semantic search over internal documentation and table metadata. Responsible AI controls, including Amazon Bedrock Guardrails and SELECT-only query enforcement at the tool level, enhance security and prevent unauthorized data modifications.
Simplified Dataset Onboarding
Dataset onboarding is drastically simplified through a metadata-driven process. Adding a new dataset requires only a JSON metadata file detailing the table’s schema, business context, and column-level semantics. This file becomes the single source of truth for the AI agent. Fields like business_description_for_the_table can include engine-specific query hints, while column_details provide natural language descriptions and sample values to improve SQL generation accuracy. The pipeline processes these metadata files, generates vector embeddings with Amazon Titan Text Embeddings, and indexes them into Amazon Bedrock Knowledge Bases, making the table discoverable by all connected agents without manual re-indexing or server restarts.
For enhanced accuracy, datasets can be associated with a business context, narrowing the search scope for the AI agent. This context mapping reduces latency and improves SQL accuracy. When a user query is submitted, the agent enriches the request with business context, converts the query into a vector embedding, searches the metadata catalog, and selects the most relevant table. It then generates an optimized SQL statement, validates and auto-corrects it for compatibility with the target engine like Amazon Athena, executes the query, and synthesizes the results into a natural language response.
Security and Scalability
The architecture emphasizes layered security, with the MCP server operating in private subnets and accessible only via the authenticated client layer. Amazon Cognito validates user identity, and tool invocations are logged to Amazon CloudWatch for a unified audit trail. This design supports multi-tenancy, allowing multiple agents to share the same tools and access controls through the standardized /mcp endpoint.
The serverless infrastructure ensures elastic scaling. Amazon ECS Fargate manages compute without requiring instance management. Downstream services like Athena and Amazon Bedrock scale automatically, enabling the system to handle varying query loads seamlessly.
This architecture supports various use cases, including enterprise knowledge assistants that combine data lake queries with document search, and internal analytics copilots for business users. The standardized endpoint also facilitates multi-agent workflows, where specialized agents can share the same MCP server and access the full tool suite.
❤️ Support Independent Journalism
Your contribution keeps our reporting free, fearless, and accessible to everyone.
Or make a one-time donation
Secure via Razorpay • 12 monthly payments • Cancel anytime before next cycle
(We don't allow anyone to copy content. For Copyright or Use of Content related questions, visit here.)
The Chenab Times News Desk
