Amazon Web Services (AWS) today added a context service to its portfolio that promises to improve the accuracy of artificial intelligence (AI) workflows in a way that also serves to reduce the total number of tokens that might need to be consumed.
At the same time, AWS is making available an AgentCore Harness to its managed AgentCore service for building and deploying AI agents that enable organizations to create a working agent with three calls to an application programming interface (API). That harness makes use of micro virtual machines to safely deploy an AI agent in a way that decouples it from any specific large language model (LLM).
Announced at the AWS New York Summit, the AWS Context service is based on a knowledge graph that catalogs datasets, dashboards, and metadata. Based on the same core knowledge graph that AWS previously embedded into the Amazon Quick AI assistant, the AWS Context Service extends the reach of a knowledge graph across multiple system relationships and business rules via integrations with other AWS services, including AWS Glue Data Catalog, Amazon SageMaker Unified Studio and AWS Lake Formation.
AWS Context publishes all key metadata from structured and unstructured sources in an Apache Iceberg format that is accessed via Amazon S3 Tables that can then be queried. The AWS Context service also connects to third-party catalogs via APIs and Model Context Protocol (MCP) servers and tools.
AWS Context also makes every query identity-aware. Each call is designed to inherit the identity and Lake Formation permissions, so an agent can only see and traverse data it has been authorized to access. That capability in turn creates an audit trail that documents those interactions.
Every interaction with that graph ultimately adds additional context that any AI agent can then later invoke to improve outcomes, says Dr. Matt Wood, chief AI and technology officer for AWS. “It’s a self learning knowledge graph,” says Wood. “The AI agent now has full situational awareness.”
AWS is also extending AgentCore Policies to evaluate every agent action for risks such as prompt injection, harmful content, and sensitive data exposure.
Additionally, Agentic Web Search is now available as a fully managed tool on AgentCore, while the Amazon Bedrock Managed Knowledge Base now provides native connectors to S3, SharePoint, Confluence, and Google Drive through which it can ingest data using an agentic retriever that plans queries across knowledge bases, connects related concepts, and re-ranks for the best answer.
AWS is also previewing a business context and semantic search for AWS Glue Data Catalog to make it simpler to discover and understand data along with a set of skill assets in Glue Data Catalog that makes it simpler to persistently access data that provides additional context for agentic workflows. Organizations can now enrich their Glue tables, views, and columns, including those backed by S3 Tables, with business descriptions, glossary terms, custom metadata, and associate them with skill assets that provide additional context stored outside the catalog.
Finally, AWS announced the general availability of Amazon S3 annotations, a new way to attach rich, queryable business context directly to S3 objects to provide AI agents with additional context.
As agentic AI workflows continue to evolve, it’s becoming apparent that the platforms being relied on to store data and provide much-needed context are as critical as the underlying AI models themselves. The challenge now is finding the best way to strike the balance needed to not only improve the quality of the outputs generated but also reduce the total number of tokens consumed.
