Google’s Data Agent Kit Hits GA, Splitting Agent Guardrails Between Skills and IAM

Google has taken Data Agent Kit out of the preview it ran through the summer and made it generally available. The kit is free and open source, and it is designed to install into coding agents that Google does not own. The launch post was written by product manager Arun Nair and developer advocate Jeff Nelson.
The package has two parts. One is a set of MCP tools — Model Context Protocol servers that expose live schema and tools — connecting an agent to more than fifteen Google Data Cloud services. The other is a collection of Google-authored skills that carry best-practice guidance, such as how to optimize BigQuery SQL or design Bigtable row keys. Supported hosts include Antigravity, Claude Code, Codex, Cursor and VS Code-compatible editors; it comes preinstalled in Cloud Shell and Cloud Workstations.
The GA release adds BigQuery Graph, Bigtable and Managed Service for Apache Spark to the services the kit covers. The authors pitch it at a gap agents have by default: they write good SQL but know nothing about which tables exist, how they are partitioned, which ones a team trusts, or why last night's job failed.
The plugin repository holds 37 skills under Apache 2.0, each with frontmatter naming Google as publisher. It is versioned, with 1.0.0 cut on September 29 and 1.1.0 on October 13, so teams can pin a release.
One of those skills is accidental-data-loss-prevention. Its description opens with a stop-and-verify instruction and tells the agent it must obtain explicit user consent before anything irreversible. Examples listed include DROP TABLE, TRUNCATE, a DELETE missing a WHERE clause or using 1=1, storage deletions against production buckets, gcloud projects delete, and KMS key destruction. The procedure is to halt, explain the impact, and wait for affirmative consent.
That guardrail is advice: the agent is asked to obey it. The second control is enforcement. Asked on LinkedIn whether read-only investigation could be isolated from tools that modify infrastructure, Nelson said it is handled in IAM rather than per agent — by running the agent as a read-only service account through impersonation, or by using an IAM deny policy that blocks any MCP tool where tool.isReadOnly is false. The documented condition is api.getAttribute('mcp.googleapis.com/tool.isReadOnly', false) == false; it denies the mcp.googleapis.com/tools.call permission and can apply at organization, folder or project level. Companion attributes cover resource.service and tool.name.
Three limits accompany that approach. A blocked tool still appears in tools/list, so the agent sees capabilities it cannot use and only finds out on invocation. The controls cover Google Cloud MCP servers and nothing else. And resource.service and tool.name are gcloud only, absent from the console.
Practitioners responded with both interest and caveats. Amel Habira, an AI solutions architect, highlighted the short-lived credentials, the read-only attribute and a warning before agents execute notebook cells, and suggested combining that with Iceberg branching and time travel to get containment at both the access layer and the data layer. Ankit Dongre of viaSocket argued the harder boundary is permission scope plus repeat-safe execution: what happens when an agent retries a write after a timeout, or when the schema changes mid-run. IAM and dry runs cover authorization and preflight, he said, but do not reveal whether a timed-out write committed; he called for tool results carrying an idempotency key with commit status and rollbackability. Data Agent Kit's tool results do not carry that today.
Microsoft takes a different route: its Agent 365 integration with Azure API Management, in public preview, keeps governance intent in a Microsoft 365 admin surface and enforces it at a gateway, where per-tool blocking needs a preview tier and the application key is gateway-scoped, reaching every model and tool. Google instead authorizes per tool in the IAM language customers already use and propagates the user's identity so row and column security applies. Timon Knapp of Conductral described a cautious rollout: start with read access and suggestions, let a human confirm anything that changes a schedule or reruns a job, and keep a clean audit trail so the agent can be given more rights later.
For developers, the practical distinction is between a polite request and a hard stop. Skills are prompts in a repository; they can be forked, edited or ignored by a model under pressure. An IAM deny policy is evaluated by Google Cloud when the call is made. This suggests teams should treat the deny policy as the real security boundary and the skills as UX that reduces friction, since a deny produces an error the agent must handle rather than a conversation it may skip. A likely trade-off: an agent that discovers blocked tools only at invocation may flail or retry needlessly, so allowlisting tools the agent can see may pair well with deny rules. The retry-safety gap is the open engineering problem; idempotency keys in tool results would be a natural next step, though nothing in the release indicates one is planned.
Data Agent Kit costs nothing beyond the Google Cloud services an agent uses, and the skills are on GitHub and forkable.
Why it matters: teams wiring coding agents into production data platforms get an authorization model built on familiar IAM primitives, plus a clear reminder that prompt-level guardrails are not enforcement. Platform and security engineers should plan identity impersonation and deny policies first, and treat repeat-safe writes as a separate problem the kit does not yet solve.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Google has made Data Agent Kit generally available, a free set of MCP tools and 37 open-source skills that connect coding agents to more than fifteen Google Data Cloud services. Two kinds of guardrail ship with it, and practitioners need to tell them apart: one is a skill asking the agent to stop, the other an IAM deny policy that refuses the call. By Steef-Jan Wiggers