What changed
Google Cloud made Data Agent Kit generally available on October 1. It packages two pieces for the coding agent you already use: MCP tools that connect to more than 15 Google Data Cloud services, and open-source skills that teach data-specific practices such as optimized BigQuery SQL, Bigtable row-key design, and pipeline construction.
The GA release adds support for BigQuery Graph, Bigtable, and Managed Service for Apache Spark, alongside connections to analytics, operational databases, lakehouse catalogs, and orchestration tools. The same layer can be installed in VS Code-compatible editors or added to Antigravity, Claude Code, and Codex CLI.
Why it matters
The meaningful product shift is not another chat interface. It is the movement of data context into the agent runtime. Instead of pasting schemas, trusted-table guidance, and failed-job logs into a prompt, a builder can ask an agent to discover the right assets, run a query, inspect execution details, and turn working logic into a scheduled pipeline.
That also gives the feature a more credible enterprise boundary: actions run with the user's IAM permissions, while Google says teams can govern MCP access with IAM, Model Armor, VPC Service Controls, and principal access boundaries. The kit is free, but the underlying Google Cloud services still incur their normal usage costs.
The product takeaway
For builders, Data Agent Kit is a useful pattern for making agents dependable: package domain knowledge as reusable skills, expose live systems through narrowly scoped tools, and keep authorization at the tool boundary. The portability is mainly in the agent surface—one package can meet Codex, Claude Code, Cursor, and other hosts—while the data plane remains Google Cloud-specific.
A sensible starting point is read-heavy workflows with dry runs and explicit approval before writes. That lets the agent prove it can find the right data and explain its plan before it is allowed to create tables, change schemas, or schedule production jobs.