The policies are enforced through a policy enforcement point, currently Agent Gateway, which must carry all of the agent's egress traffic to its model. When the model returns a tool call, the gateway sends it to the policy engine, where an LLM judge compares it against the conversation, the agent's system instructions, the tool descriptions and the written policies. A call runs only if it passes both the user intent check and the rules; otherwise the engine swaps the response for a configurable denial message.
Each policy is a natural-language constraint attached to an agent, and it can cover all of the agent's tools or a single tool, including its parameters. The announcement's example is an earnings-disclosure rule that keeps unreleased financial results within company addresses until the public release. Cloud Logging records each verdict and the judge's rationale, and text-only responses are not evaluated because they take no action.
Why it matters if you build agents
Developers can add a check on every proposed tool call without writing policy code, but LLM verdicts can vary, so Google recommends running policies in dry run mode before enforcing them. The judge only sees the conversation, the system instructions and the policies, so facts a policy depends on, such as an earnings date, belong in the agent's system instructions. Semantic governance policies sit alongside permissions and Model Armor rather than replacing them.
Get stories like this every weekday
One short email with the day's agent tooling news.