---
title: "Check what your AI agents do before they do it, with Semantic governance policies"
description: "Google Cloud introduced Semantic governance policies in preview for the Gemini platform, enabling enterprises to govern AI agent tool calls before they execute. The policies check each proposed tool call against user-written plain-language rules and the original user intent, preventing actions that violate business rules or result from prompt injection. The enforcement sits outside the agent code via a policy enforcement point such as Agent Gateway, which intercepts model responses containing tool calls. An LLM judge evaluates each call against the conversation, agent instructions, and policies, either allowing it or replacing it with a denial message. The feature supports models from Gemini, Anthropic, and OpenAI-compatible APIs, and logs every verdict and rationale to Cloud Logging for auditability."
---

# Check what your AI agents do before they do it, with Semantic governance policies

[Google](https://yomu.fyi/company/google) · Deepti Sehrawat · Nov 20, 2026

**Type:** Announcement

## Summary

Google Cloud introduced Semantic governance policies in preview for the Gemini platform, enabling enterprises to govern AI agent tool calls before they execute. The policies check each proposed tool call against user-written plain-language rules and the original user intent, preventing actions that violate business rules or result from prompt injection. The enforcement sits outside the agent code via a policy enforcement point such as Agent Gateway, which intercepts model responses containing tool calls. An LLM judge evaluates each call against the conversation, agent instructions, and policies, either allowing it or replacing it with a denial message. The feature supports models from Gemini, Anthropic, and OpenAI-compatible APIs, and logs every verdict and rationale to Cloud Logging for auditability.

## Context

Google Cloud announced a preview of Semantic governance policies for the Gemini platform, designed to prevent AI agents from executing tool calls that violate enterprise rules or deviate from user intent. The feature addresses risks such as indirect prompt injection and social engineering that traditional permissions and system instructions cannot block.

## Approach / What changed

Semantic governance policies are enforced through a policy enforcement point (Agent Gateway) that intercepts each tool call an agent’s model proposes before it executes. The policy engine uses an LLM judge to evaluate the proposed call against the conversation history, the agent’s system instructions and tool descriptions, and user-written plain-language policies. If the call passes both the user intent check and the policy check, it runs; if it fails, the engine replaces the response with a configurable denial message. Policies can be created in the Google Cloud console or via the gcloud CLI, and the engine supports models from Gemini, Anthropic, and OpenAI-compatible APIs.

## Takeaways

- Semantic governance policies evaluate each tool call an agent proposes against both a built-in user intent check and custom plain-language policies before the call executes.
- Policies are written in natural language by legal or compliance teams, require no code or redeployment, and can apply to all tools or specific tool parameters.
- The policy engine supports models from Gemini, Anthropic, and OpenAI-compatible APIs, and logs every verdict and the judge’s rationale to Cloud Logging for auditability.

## FAQ

### How does the policy engine evaluate tool calls?

The policy engine uses an LLM judge to evaluate each proposed tool call against the conversation history, the agent's system instructions and tool descriptions, and your plain-language policies. The call must pass both a built-in user intent check and your custom policy rules before it is allowed to execute.

### Can I test policies without blocking actions?

Yes, you can run the policy engine in dry run mode, which logs what it would deny without actually blocking any calls. This allows you to review verdicts in Cloud Logging and refine your policies before switching to enforcement mode.

### Which models does the policy engine support?

The policy engine supports models accessed through the Gemini API, the Anthropic Messages API, Claude models in Gemini platform, and OpenAI-compatible Chat Completions APIs, including self-hosted servers such as vLLM.

**Tags:** [AI Agents](https://yomu.fyi/topic/ai-agents), [Gemini](https://yomu.fyi/topic/gemini), [Google Cloud](https://yomu.fyi/topic/gcp), [Governance](https://yomu.fyi/topic/governance), [Security](https://yomu.fyi/topic/security)

- Source: [Google](https://developers.googleblog.com/check-what-your-ai-agents-do-before-they-do-it-with-semantic-governance-policies)
- Source URL: https://developers.googleblog.com/check-what-your-ai-agents-do-before-they-do-it-with-semantic-governance-policies
- Ingested by Yomu: 2026-11-21T00:00:23.125Z

[Read original post](https://developers.googleblog.com/check-what-your-ai-agents-do-before-they-do-it-with-semantic-governance-policies)
