Vibitely Daily

Sunday, 22 November 2026, story 4 of 7

Google adds plain-language tool-call checks to Agent Gateway

Semantic governance policies, in Preview, use an LLM judge to test each proposed agent tool call against written rules and the user's request.

Illustration: A row of identical parcels moves along a conveyor toward a gate, where a translucent checkpoint arch lets some through and diverts one onto a side ramp.

What happened

Google has released Semantic governance policies in Preview in its Gemini platform, according to a post on the Google Developers Blog. Routed through Agent Gateway, each tool call an agent's model proposes is judged by an LLM against two things: rules written in plain language, and what the user asked for. A call must pass both to run.

If a call fails, the engine replaces the response, tool calls included, with a configurable denial message. Verdicts and the judge's rationale go to Cloud Logging. Text-only and streamed responses are not evaluated. The post says a dry run mode logs what would be denied without blocking anything.

What it means

The design puts the check outside the agent's code, which matters for prompt injection: a hidden instruction in a document can sway the agent's model, but not a separate judge. Policies written by legal or compliance staff can change without a redeploy, according to Google.

The limits are plain in Google's own post. Verdicts come from an LLM, so they may vary, and Google advises running dry mode before enforcing and keeping other controls in place. Streamed responses are skipped, so teams that stream may have gaps. The judge also depends on clear tool descriptions and on facts, such as an earnings date, being in the system instructions. Google says it will publish AgentDojo benchmark results in a later post. Until then, there is no independent evidence of how well it blocks attacks.

What to do: If your agents run behind Agent Gateway, switch on dry run mode, give every tool a clear description, and read the logged verdicts before enforcing any policy.

Source

Filed under Agents. Published .

"What it means" is analysis, not reporting. Found an error? Write to daily@vibitely.com.