AI agents move from passive analytics consumers to active BI authors as dbt 2.0's proprietary fork reshapes the data transformation stack.

The week in analytics

Two patterns defined this week. The first: AI agents are no longer just querying analytics infrastructure — they're authoring it. Basedash made its MCP server writable so agents can create charts directly, and Hex shipped a chat agent that can build and publish entire Hex projects from conversation. These aren't demos; they're production capabilities landing in the same week, signaling a step change in how the analytics creation layer is being instrumented.

The second pattern is structural: dbt 2.0 shipped. dbt Core's GA release officially splits the runtime into a proprietary dbt binary and a community dbt-oss package — the most consequential vendor move in the data transformation space since dbt Labs went enterprise. The week's other releases in the semantic layer and BI segments all land against this backdrop, because the dbt fork changes how governed definitions flow between transformation and analysis.

Leaders

Basedash shipped two interlocking releases that together constitute a category move. Models introduces a semantic layer — reusable, governed SQL definitions that named business concepts like 'active users' can be built on by the whole team, including AI. Then the writable MCP server makes those definitions actionable: AI agents running in Cursor, Claude, or any MCP client can now create real charts and dashboards in Basedash from a prompt, not just read existing ones. Most BI tools with MCP integrations are read-only. Basedash went past that line.

dbt Core released v2.0.0 GA this week — codenamed 'Benjamin Franklin' — formalizing the split between a proprietary dbt binary (with AI agent skills and closed enhancements) and dbt-oss (the Apache-licensed community fork). The implications run deep: teams that assumed dbt was purely open-source now face a choice, and every BI and analytics vendor that integrates with dbt — including several others in this sector — needs to articulate which binary they support.

Holistics connected to Snowflake semantic views and Databricks metric views as first-class sources this week, syncing governed definitions directly into the Holistics model layer. This is Holistics shifting from semantic owner to governed consumer: instead of asking teams to redefine their metrics inside Holistics, it pulls them from where the data team has already defined them. For organizations with mature Snowflake or Databricks setups, this removes the duplication that has historically made adding a BI tool painful.

Omni shipped two structural moves in the same release cycle: dbt integration on Trino-based connections (adding Trino to the Spark and BigQuery coverage already in place) and MCP server tools for managing Omni Apps. The Apps themselves reached GA — exiting preview as a first-class product, now with MCP clients able to export dashboard PNGs. The Trino addition matters because it extends governed analytics into one of the most common large-scale query engines that wasn't previously covered.

Hex gave its chat agent the ability to create and modify Hex projects from conversation — not just answer questions, but build the analytics artifact itself. Alongside this, versioned eval suites shipped so teams can publish and test AI-generated analyses before promoting them. The approval-bypass feature is an operational detail that tells a story: Hex expects agents to be creating enough projects that requiring human approval for each one becomes a bottleneck.

Wildcards

Plausible added a dedicated AI Assistants channel to its Channels report, breaking out ChatGPT, Claude, Gemini, and Perplexity as individually trackable traffic sources. This is a data product decision that reflects an observable shift: AI assistants have become measurable referrers, not just a tail category lumped into 'other.' For analytics teams tracking acquisition, this channel is now table stakes. Plausible shipping it as a named feature makes AI referral a first-class dimension.

Lightdash shipped custom chart types this week that accept natural language descriptions — users describe what they want and Lightdash generates the chart type as a reusable extension. This is a quieter move than Basedash's full agent authoring, but it changes the creation surface: custom chart development moves from writing Vega-Lite specs to describing intent. The governance model behind it — chart types as shared extensions the whole org can reuse — mirrors the semantic layer pattern Holistics and Basedash also shipped this week.

Themes that compounded

• Agent-authored analytics shipped at Basedash (writable MCP) and Hex (chat agent builds projects) in the same week — this pattern has now moved past the single-product experiment stage.
• The semantic layer is becoming the access control boundary for AI agent analytics: Basedash's Models and Holistics's warehouse-native sync both frame governed definitions as the gating layer for what agents can touch.
• dbt Core's proprietary fork will pressure every BI vendor in this sector to clarify their dbt-oss vs dbt-proprietary integration position over the coming weeks.
• AI-as-traffic-source is now a measurable analytics dimension: Plausible and Microsoft Clarity both shipped AI-specific analytics features this week.

Watch this week

The dbt 2.0 fork will surface downstream: watch for integration announcements from BI vendors clarifying which binary they're building on, and for data teams to start asking whether dbt-oss covers their use case or whether the proprietary binary features they want require a paid relationship with dbt Labs. On the agent authoring front, Basedash's writable MCP is the most direct test of whether customers actually want AI writing production dashboards — adoption signals over the next two weeks will be worth watching.