What changed
BTN previously covered Google Cloud's August 31 Data Agent Kit launch as an open-source MCP/skills layer for agentic data pipelines. On September 30, 2026 Google declared it generally available, and its October 9 Data Cloud digest reaffirmed the rollout. GA expands to more than 15 Google Data Cloud services and adds BigQuery Graph, Bigtable and Managed Service for Apache Spark alongside existing Spanner, AlloyDB, Cloud SQL, BigQuery, dbt/Dataform and Airflow orchestration. VS Code-compatible IDE extensions and plugins for Claude Code, Codex and Antigravity now provide shared project context, SQL/notebook editing, data browsing and guided pipeline debugging. Setup can enable required APIs and configure matching MCP servers and skills automatically. Google says the kit costs nothing separately, but users pay normal cloud-service charges.
Why it matters
The development turns an early integration kit into a more complete, generally available developer workflow: a coding agent can discover schemas, author queries, manage data resources and diagnose failed jobs without constantly switching to cloud consoles. For teams already using Google Cloud, this can reduce setup friction and make data engineering more reproducible in an IDE. GA is not evidence that agents are safe to operate autonomously: all operations run under the user's or impersonated service account's IAM permissions, and costly or destructive queries still require normal controls, dry runs and review. The open-source skills are portable across agent clients, not across the underlying cloud products.
The agent interface is moving into the developer’s existing environment
Google is packaging data-product knowledge and actions for IDE and CLI agents instead of requiring builders to shift into a separate conversational console. The launch explicitly describes use from VS Code, Claude Code and Codex, giving teams a way to keep repository context, code review and existing developer workflow around data-pipeline changes.
Orchestration Pipelines becomes the execution model behind the conversation
The kit connects agent instructions to Google Cloud’s Orchestration Pipelines framework. That means the output is intended to become concrete pipeline configuration and workflow artifacts rather than remaining a prose answer. Builders can use the agent to reason about pipeline construction while preserving an inspectable deployment surface.
Open source makes the integration layer inspectable
Google published Data Agent Kit on GitHub with product extensions, MCP servers and related agent tooling. Teams can inspect how capabilities are exposed, adapt extensions and evaluate exactly what an agent is allowed to call. Open sourcing the kit does not make the underlying Google Cloud services portable, but it reduces dependence on a single proprietary agent client.
Agentic data work still needs ordinary infrastructure controls
A coding agent that can create or change data pipelines can also create expensive, destructive or over-privileged infrastructure. Teams should keep service-account scope, environment separation, code review, schema validation, dry runs and deployment policy around the generated work instead of treating the agent interface as a replacement for those controls.
September GA broadens the supported data surface
The September 30 release adds BigQuery Graph authoring/query skills, Bigtable schema and row-key guidance, and Managed Service for Apache Spark sessions over Iceberg Lakehouse data. Data Agent Kit now connects more than 15 Google Data Cloud services, extending the original pipeline-oriented release into database and analytics workflows. Google reports that the extension can preview graph relationships and propose changes for approval; these are product capabilities, not independently verified correctness guarantees.
Developer installation, governance and pricing are now explicit
The kit is available through VS Code-compatible extensions, Claude Code/Codex/Antigravity plugins and preinstalled Cloud Shell/Cloud Workstations. Sign-in can automatically enable APIs and configure MCP servers. Operations run with user IAM or impersonated service-account authority, with optional Model Armor, VPC Service Controls and Principal Access Boundary protections. The kit itself has no separate fee; BigQuery, Spark, storage and other invoked cloud resources incur their usual charges. Operators should restrict privileges, validate queries and keep approvals around expensive or destructive actions.