AWS Enters the 'Context Layer' Race With a Self-Learning Knowledge Graph for AI Agents

The generative AI boom has produced a fleet of highly capable reasoning engines, but deploying them inside a large enterprise often reveals a glaring limitation: they have no idea what a company's data actually means. An AI agent might know how to write flawless SQL, but it does not inherently know that the "revenue_Q3" column in a legacy database excludes European sales, or that a specific storage bucket contains deprecated customer records.

Until now, solving this "context gap" required data engineering teams to manually build bespoke Retrieval-Augmented Generation (RAG) pipelines. Engineers had to hand-curate metadata, define business relationships, and constantly update the system as database schemas changed. It is a brittle, labor-intensive process that scales poorly as organizations attempt to deploy hundreds of specialized AI agents across different departments.[1]

Amazon Web Services (AWS) is making a direct play to eliminate that manual labor. At the AWS Summit in New York City on Wednesday, the cloud computing giant unveiled a suite of tools designed to automate how AI agents understand enterprise data, centered around a new service called AWS Context.[1][5]

AWS Context functions as a dynamic, self-learning knowledge graph. Rather than forcing human engineers to map every relationship, the service automatically scans an organization's existing data stores—across data lakes, warehouses, and databases—and infers how different tables, columns, and business rules relate to one another.[1][4]

"You want to trust the decisions made by your AI agents, but that can't happen until agents have context," Mai-Lan Tomsen Bukovec, AWS vice president of technology for data analytics, wrote in a blog post detailing the launch. The goal is to provide a governed, centralized layer that gives agents the situational awareness required to act autonomously without hallucinating.[2][3]

The most significant architectural shift in AWS Context is its self-learning mechanism. The knowledge graph does not just map data statically; it observes how AI agents interact with it in real-time, effectively crowdsourcing intelligence from the AI models themselves.[1][3]

When an agent queries the graph to solve a problem, AWS Context monitors which data sources produce correct results, which join paths the agent relies on, and which curated business rules are successfully applied. It then ranks those sources by actual usage, continuously refining the enterprise ontology.[3][4]

"Your agents now get smarter without you having to rebuild anything from scratch," Swami Sivasubramanian, vice president of Agentic AI at AWS, explained during his keynote address. If one customer support agent discovers the correct path to resolve a schema ambiguity while pulling a user's purchase history, the graph updates automatically. The next agent to face a similar query will instantly know the optimal route, bypassing the need for human re-curation.[1][4]

To feed this centralized graph, AWS introduced two supporting features that push context closer to where the data actually lives. The first, Amazon S3 Annotations, is now generally available. It allows organizations to attach rich, queryable business context directly to individual objects stored in Amazon's ubiquitous S3 storage service.[1][3]

The second feature, currently in preview, introduces skill assets to the AWS Glue Data Catalog. This allows data producers to attach domain-specific knowledge—such as team runbooks, query patterns, and usage rules—directly at the catalog layer. AWS Context then synthesizes the S3 annotations and the Glue metadata into a single, cohesive map that agents can query at runtime.[1][3]

Industry analysts note that this approach addresses a critical bottleneck in enterprise AI adoption. Jake Dolezal, lead data engineer at McKnight Consulting Group, highlighted that AWS Context represents a "meaningful shift from each team building its own pipeline to one governed context layer the whole organization draws from."[2]

By extending an existing identity model, AWS is also addressing the security nightmares associated with autonomous agents. AWS Context makes all queries identity-aware. An AI agent can only see, search, and operationalize information that it is explicitly authorized to access, preventing a rogue agent from surfacing sensitive HR or financial data to an unauthorized user.[2][4]

The launch places AWS squarely in the middle of a fiercely contested architectural battle. The "context layer" has rapidly become the most valuable new real estate in the enterprise software stack. Snowflake recently announced its Horizon Context and Cortex Sense services, while Microsoft offers a semantic ontology through its Fabric IQ platform.[1]

Independent vendors are also racing to solve the same problem. Vector database provider Pinecone offers a Nexus context service that compiles enterprise data into task-specific artifacts, and Redis has developed its own retrieval-optimized context platform.[1]

AWS's structural argument to enterprises is one of zero-integration friction. For companies already running their data lakes on Amazon S3, AWS Glue, and Lake Formation, AWS Context requires no data movement. The metadata is simply published into the open Apache Iceberg format, allowing customers to query their context with tools like Amazon Athena or Apache Spark.[1][3]

As the industry transitions from chat-based copilots to proactive, autonomous workflow agents, the underlying infrastructure must evolve. The era of siloed data and manual RAG pipelines is giving way to automated, self-healing knowledge graphs—ensuring that the AI agents of tomorrow actually understand the businesses they are trying to help.