Google Cloud unveils Data Agent Kit for developers

Google Cloud has introduced Data Agent Kit, a preview product for data developers that lets users run data workflows from their development environments.

The offering combines Model Context Protocol servers with agent skills, allowing software agents to query different data systems and return results inside tools such as code editors and command-line interfaces. The aim is to reduce the need for data practitioners to move between separate consoles, databases and storage systems when answering open-ended business questions.

Data Agent Kit is available as an extension for VS Code forks, including Antigravity IDE and Cursor, and as a plugin for tools including Antigravity 2.0, Antigravity CLI, Claude Code and Codex. The system relies on Model Context Protocol, an open standard for connecting agents to tools, databases and cloud infrastructure, and on skills written as markdown files that guide an agent in interacting with a specific technology stack.

Rather than producing SQL for a user to copy into another interface, the toolkit lets an agent execute queries and inspect results on the user's behalf. The process still requires approval for tool use, and the development environment can pause to request permission before an agent runs actions such as a read-only SQL query.

Cross-system analysis

Google Cloud illustrated the software with a retail analysis example focused on a drop in average order value while revenue stayed flat. In that scenario, analytical sales data sat in BigQuery, customer records were stored in a Cloud SQL PostgreSQL database and marketing campaign rules were held as raw JSON files in Cloud Storage.

The workflow began with a natural language prompt in the integrated development environment asking the agent to calculate monthly average order value from orders and order-items tables. After permissions were granted, the agent queried BigQuery and found that average order value had been about $110 from August to December before falling to about $103 in January.

A follow-up prompt asked the agent to break down January's average order value by order type. The results showed that online and offline orders remained close to $110, while a new B2B-Wholesale channel appeared in January with an average order value of about $75.

The analysis then moved to Cloud SQL, where the agent examined customer records linked to the wholesale orders. In Google Cloud's example, all 100 wholesale accounts were newly created business entities added within the previous 30 days.

The investigation next moved to campaign files in Cloud Storage. There, the agent traced a promo code used on 92% of B2B orders to a marketing campaign offering a 25% discount, explaining the lower blended average order value without a fall in overall revenue.

From analysis to models

Google Cloud also presented the toolkit as a way to turn one-off investigations into repeatable data projects. In its example, the user asked the agent to build a dbt project that joined BigQuery staging models with Cloud SQL customer and pet profile attributes, added a uniqueness test on order_id and ran dbt build.

The first build failed because customers could own more than one pet, causing duplicate order rows when those profiles were attached directly to each order. The agent reviewed the terminal output, rewrote the dbt logic and reran the build until the test passed.

That example highlights both the appeal and the limits of the approach. The software may cut the time spent writing repetitive queries across multiple systems, but users still need to review generated code and apply data quality checks to the output.

Google Cloud placed particular emphasis on visibility into agent activity. Users can inspect the execution trail in the development environment, including individual Model Context Protocol tool calls and the raw SQL sent to BigQuery, giving teams a way to audit what the agent has done.

The launch comes as large technology groups try to embed AI agents into developer workflows and data operations rather than limit them to chat-based assistance. For cloud providers, the opportunity lies in linking natural language interfaces to databases, object stores, transformation frameworks and operational systems that businesses already use.

Google Cloud is positioning Data Agent Kit around the practical work of investigation: tracing a business anomaly across warehouse data, production databases and unstructured files, then turning the result into a reusable model within the same working session. Its preview status suggests the company is still testing that workflow with developers before a broader rollout.

Google Cloud said users can inspect generated queries because "It's important to keep an eye on generated code, though reading a query can take much less time than writing one against schemas you're unfamiliar with."