Rung 8 — Golden layer + semantic layer for AI agents

One line: Give machines the same trustworthy definitions humans fought for, so an agent's answer is the company's answer.

Concept. You leave Rung 6 or 7 because AI agents and LLM applications are being pointed at the data and either hallucinating joins or producing three different revenue numbers. A golden layer is the curated, certified set of entities and metrics the business stands behind. A semantic layer sits on top and defines meaning — what "active customer" is, how "net revenue" is computed, which table is authoritative — as metadata that both BI tools and agents query through, instead of writing raw SQL. The result: an agent asked "how did the Boston store do last week?" resolves "Boston store", "last week", and "did" the same way the CFO's dashboard does. Governance stops being a document and becomes an API.

Signs you are here. Business: leadership asks questions in natural language and trusts the answers; an "AI answers" product is customer-facing; metric definitions are versioned and owned. Technical: a metrics/semantic layer (dbt Semantic Layer, Cube, LookML, AtScale) in front of the warehouse; certified datasets flagged in the catalog; agents using tool-calls or MCP servers that expose metrics not tables; evaluation suites for agent answers; RAG over documents grounded with the same entity IDs.

The decision. Semantic layer product or hand-rolled metric definitions in dbt? A product when many consumers (BI, agents, apps) must agree; dbt metrics alone when the consumer is one BI tool. Agents query through the semantic layer (safe, governed, less flexible) or generate SQL against gold (flexible, risky)? Through the semantic layer for anything customer-facing or financial; SQL generation only in analyst-assist tools with a human reviewing. How is "certified" decided — by a data council, by domain owners, by tests? Tests plus an owner's sign-off, recorded in the catalog. Which model for which job: reserve the strongest model for reasoning over metrics, a cheaper one for routine retrieval and chat.

Tools. Semantic layers: dbt Semantic Layer / MetricFlow, Cube, Looker (LookML), AtScale, Snowflake Semantic Views, Databricks Unity Catalog metrics. Agent frameworks and protocols: Model Context Protocol (MCP) servers over the semantic layer; Claude API with tool use; LangGraph; Bedrock Agents, Vertex AI Agents. Vector and retrieval: pgvector, Pinecone, Weaviate, OpenSearch. Evaluation and observability: LangSmith, Braintrust, Arize Phoenix, promptfoo. Governance: catalog certification flags, lineage into the semantic layer.

Next trigger. There is no Rung 9 yet. The signal that the ladder has served its purpose: new questions can be answered by defining a metric, not by building a pipeline. The next moves are lateral — more domains, more clouds, more agents — not upward.

Threads. Security: agents are a new principal — scope their identity to the semantic layer, never to raw tables; prompt-injection through retrieved documents is a data-security problem, treat it as one; log every question and answer. Compliance: explainability — an answer must cite the metric definition and data as of a timestamp; GDPR applies to what the agent memorises; HIPAA de-identification must happen before the vector store. Data modeling: entity resolution (one customer ID everywhere) is the hard prerequisite; metrics defined once with dimensions, grain, and filters; conformed time. Cost: LLM tokens per question become a data-platform line item — cache, route cheap questions to cheap models, cap per user; semantic-layer compute is bursty, size for peaks.

[k3r war-story slot: the agent that answered confidently from the wrong table]