September 2026 Summaries

Snowflake CoCo, formerly Cortex Code, can connect to Cube through the Model Context Protocol or Cube CLI to query governed data, analyze semantic models, and create shareable workbooks and dashboards. Teams can continue defining metrics in Snowflake using Semantic Views, import those definitions into Cube, review and deploy generated models, and update them when source definitions change. Cube supplies the governed data model, query execution, permissions, and content storage, while CoCo directs investigation through tools that discover metrics and dimensions, run queries, and save reports. Connection options include configuring Cube as an MCP server with OAuth or installing Cube’s CLI and open-source agent skills, which guide CoCo through model exploration, query validation, and dashboard creation. The demonstrated workflow uses an e-commerce model to build a revenue and orders dashboard that teammates can explore by adjusting date ranges, time granularity, departments, categories, or brands.

Cube Store is a distributed columnar serving engine designed to answer analytical queries from pre-aggregated data efficiently, using aggregate awareness in Cube’s semantic layer to match queries with eligible rollups or fall back to source data when necessary. Built in Rust with Apache Arrow, DataFusion, Parquet, RocksDB, and object storage, it separates routing, metadata, distributed worker execution, refresh operations, and local data caching. Its performance approach relies on sorted indexes, partition pruning, Parquet statistics, local worker copies, partial and streaming aggregation, and incremental refreshes to reduce data reads, intermediate processing, and coordination overhead for recurring query patterns. In a study of 11 adapted TPC-H queries across source scales ranging from about 6 million to 6 billion line items, 43 reported query-and-scale combinations had median latencies of 173 to 620 milliseconds and p90 latencies of 291 to 879 milliseconds, although the results did not evaluate concurrent throughput, refresh costs, or the separate impact of individual optimizations.

Cube’s shift from a universal semantic layer to agentic analytics has centered on building an “agentic analytics harness” that governs model context, tools, memory, validation, and permissions to make AI behavior reliable in production. The company found that SQL gives agents the flexibility to explore unanticipated questions, while Semantic SQL preserves governed metric definitions, joins, aggregations, and access controls rather than exposing raw warehouse logic. Reliable tools require precise descriptions, validation, and recovery-oriented error messages so models can understand failures, partial results, and next steps without guessing. Because context is limited, Cube truncates and paginates query results, separates specialized tasks into different agents, and stores persistent work in durable artifacts such as version-controlled data model changes. The semantic layer supplies executable business definitions and policies, but agents also need distinct forms of contextual knowledge, including documentation, certified queries, connectors, and scoped memory. Rather than preloading an entire potentially outdated model, Cube searches the semantic model on demand, filters results by user access, and uses metadata and lineage to help agents select among similar measures.