AWS COA Open-Sourced: Semantic Layer Evolves into Agent Runtime Context Layer

01 Semantic Layer First Solved "How to Calculate"; Agents Now Ask "How Does This Business World Operate"

Traditional semantic layers centralized metrics, dimensions, relationships, and business terms to resolve "same name, different meaning" and "same metric, different numbers" across reports. BI tools could then share a single business definition. Agents push this further: models can generate SQL and call tools, but without governed context they must guess which entity a "customer" maps to, what the contract-account relationship is, which revenue metric finance approves, and what conditions trigger approvals. Stronger query capabilities amplify the risk of ungoverned guesses entering automated execution chains.

Apache Ossie (Incubating) now includes ai_context in its open semantic model, allowing the same model to carry metadata for AI tools. AWS goes further with its Prescriptive Guidance "Semantic Layer for Agentic AI," placing ontology, symbolic reasoning, virtual knowledge graphs, metric definitions, and MCP/A2A interfaces inside one semantic system. The semantic layer's center of gravity is shifting: not just making every report compute the same number, but making different agents understand the same business world.

02 RAG Finds Documents, but Complex Business Questions Need Relationships, Rules, and Governed Metrics

RAG excels at "answer exists in a document" queries: retrieve relevant passages, let the model compose an answer. Many enterprise decisions, however, span structured data, entity relationships, and business rules simultaneously — document retrieval alone cannot complete the full judgment chain.

AWS Builder Center illustrates this with a supply-chain example: a distribution center faces a supplier delay and must find an alternative supplier in the same region and category, with a risk score below a threshold, that does not violate blocking rules. The answer is scattered across suppliers, distribution centers, transport routes, risk scores, and procurement rules. The system must know which entities connect, which relationships to traverse, and which conditions must hold simultaneously.

This maps to three distinct semantic needs: Governed Metrics provide deterministic KPI definitions; Semantic Model supplies stable business definitions and data relationships; Ontology formalizes entities, relationships, rules, and constraints. RAG, Semantic Layer, and Ontology are not replacements but complementary layers in a single context chain.

Figure 1: AWS Knowledge Graph / GraphRAG reference architecture for cross-entity relationship queries (not COA product architecture). Source: AWS Architecture Center

03 COA Turns Semantic Assets into Agent-Callable Runtime Capabilities

COA's workflow splits into three phases:

Scan : connects databases, data catalogs, and documents to build a foundation for business metadata and entity extraction.

Model : uses AI to draft an ontology, then human experts review, edit, and approve it before publishing to a knowledge graph; governed metrics can be defined separately.

Serve : exposes these capabilities to agents via MCP, REST, and SPARQL endpoints.

The key advance is not the knowledge graph itself but unifying two previously scattered asset types into one service chain: deterministic business metrics and formalized entities, relationships, rules, and constraints. Agents no longer need to re-infer business logic from raw schemas and prompts on every call; they can invoke pre-governed semantics first.

COA adopts layered resolution: questions answerable by a governed metric take a deterministic path; those needing structured relationship queries route to the ontology/virtual knowledge graph; complex cases fall back to agentic reasoning. Official documentation also designs Cedar authorization and SQL Firewall — semantics define "what the business means," while authorization and execution controls govern "who can access and what operations may run."

MCP plays a second critical role: historically semantic assets were locked inside a specific BI tool or data platform. A standardized interface lets the same metrics, entity relationships, and business rules serve multiple agents simultaneously. Models and agent frameworks can be swapped without rebuilding the enterprise's accumulated business semantics.

Figure 2: AWS MCP Server production deployment reference architecture for agent access via unified interface (not COA-specific). Source: AWS official materials

04 The Semantic Layer's Next Stop Is the Context Layer

When AWS announced AWS Context in June, it used the term "shared, governed context layer" — automatically mapping enterprise data relationships, organizing business rules, domain knowledge, and governed data relationships into a knowledge graph, and serving it to agents at runtime. On July 31, with COA's release, AWS stated that COA's user-defined ontology capability will become a native managed feature of AWS Context.

This roadmap expands the traditional semantic layer's boundary. The old center was metrics, dimensions, joins, and business terms; agent-ready context also requires complex entity networks, rules and constraints, knowledge provenance, identity permissions, and traceable resolution processes. Semantic Model, Ontology, and Context Layer are not synonyms, but they are converging around a single goal: turning enterprise business knowledge into machine-readable, governable, reusable context.

This also explains why stronger models make semantic foundations more critical. Models can write more complex SQL, call more tools, and execute longer workflows, but they cannot decide for the enterprise "which metric is the official definition," "which relationship reflects the real business," or "which rule must never be broken." Once these are not stably modeled, errors amplify faster with automation.

Models handle reasoning; the semantic layer turns the enterprise world into reason-able objects.

The agent-era semantic layer therefore does not stop at the "metrics layer." It is combining with ontology, knowledge graphs, MCP, and governance mechanisms to become a business context layer between models and enterprise data. The semantic layer is not being replaced by ontology; it is evolving from the BI-era business semantic model into the agent-era business context infrastructure.

References

AWS What's New | COA Release (2026-07-31)

AWS COA Docs | Context Ontology Accelerator

AWS Builder Center | Beyond RAG (2026-08-15)

AWS Prescriptive Guidance | Semantic Layer for Agentic AI

AWS ML Blog | AWS Context (2026-06-17)

Apache Ossie (Incubating) | Official Documentation