Evidence Core gives agent-built BI a code-review boundary

Evidence has open-sourced the framework underneath its hosted business-intelligence product as Evidence Core, betting that analytics artifacts should be readable, editable and testable by both coding agents and humans. The Aug. 26 release separates the product into an open-source Core framework and the hosted Evidence Studio platform built on top of it. Evidence’s launch post says Core connects directly to a warehouse, supports reusable metric definitions, validates its new declarative syntax statically and can be self-hosted.

That combination matters more than the “AI builds dashboards” pitch. It creates a review boundary between an agent proposing analytics and an organization publishing them.

Put definitions where reviewers can see them

Evidence says a metric such as revenue can be defined once in YAML and reused across components. If the definition changes, teams update one artifact instead of searching through every chart and table. For agent-built analytics, that makes a consequential business-logic change visible as code rather than burying it inside a generated visualization.

Core’s new syntax is also statically validatable. The launch post says an agent can receive a specific error for invalid markup without first running the project and inferring what failed. That does not prove the SQL is correct or the metric matches the business definition, but it does remove a class of structural failures before runtime.

The practical gate should therefore be: let the agent generate or edit the project, run validation, inspect the metric and SQL diff, and only then merge. Evidence’s CLI documentation describes a local development server, query and metadata commands, code validation and a Git-based workflow in which a pull request gets a preview link and merging releases the change.

Know where Core ends and Studio begins

Evidence’s product documentation describes Core as an MIT-licensed framework using SQL and Markdown, with direct warehouse connectors and self-hosting options. The hosted Studio layer adds collaborative editing, preview deployments, an AI report-building agent, access rules, page-level access control and scheduled reports.

That split is an operating decision, not a packaging footnote. A team can use Core to keep analytics artifacts portable and deploy them on its own infrastructure. Teams that need hosted collaboration and finer publishing controls should evaluate Studio separately rather than assuming every governance feature ships in the open-source layer.

Evidence also says Core’s CLI includes a migration command for its previous open-source project. Existing users should test that migration in a branch and compare rendered output, queries and metric definitions before replacing a production build.

The deployment test

Before allowing a coding agent to modify a live analytics project, create one deliberately invalid component and one deliberately changed metric. Confirm that validation catches the structural error, the pull-request diff exposes the metric change, the preview uses the intended warehouse connection and the production release happens only after approval.

Evidence Core does not make generated analysis correct. It does something narrower and useful: it turns more of the analytics surface into artifacts that can be validated, reviewed and rejected before users rely on them.