Stone columns at the front of a government building.

Google now checks each agent action against your written rules first

2 days before an earnings release, an employee asks a finance agent to send the early results to an investor and to copy a personal email account. The agent holds the file access and the send access, so the request sits inside its permissions. On November 20, 2026, Google put a check in front of that call.

Direct answer: Semantic governance policies, in Preview in the Gemini platform, judge each tool call an agent proposes before the call runs. The call has to pass 2 checks. It must follow rules your firm writes in plain language, and it must match what the user actually asked for. A failed call never runs, and every verdict lands in a log with the judge’s reasons.

How does the check sit between the model and the tool?

The check sits outside the agent’s own code, at a policy enforcement point. Today that point is Agent Gateway, and the agent’s traffic has to route through it. When the model proposes a tool call, the gateway sends the proposal to a policy engine before the answer reaches the agent.

Advertisement

A model acting as judge reads the proposal against the conversation so far, the agent’s system instructions and tool descriptions, and your written policies. A pass lets the call through unchanged. A fail replaces the call with a denial message your firm can set in its own words.

The worked example shows the split. An email with the results to the company controller passes, because the employee asked for it and the policy allows that address. The email to the investor fails on the policy, and the whole call dies. A delete call planted inside a planning document fails on the intent check, because no person asked for it.

Who writes the rule, and who answers for the verdict?

Legal, compliance and business teams write the policies in plain language, with no code change and no redeploy. A policy attaches to an agent, and it can cover every tool or 1 named tool. The authority writing in When an agent buys for you, set the authority before the first order now has a machine reader. A rule nobody wrote down still binds nobody, and this check enforces only the rules a firm bothered to write.

The accountability line stays with the firm. The judge is a model, and the vendor states plainly that verdicts may vary. 2 classes of response sit outside the check entirely. Replies that carry only text take no action and get no review, and streamed responses are never evaluated. The vendor recommends a dry run first, where the engine logs what it would deny and blocks nothing, and it presents the feature as 1 layer beside access controls and screening tools. The value of the line depends on the quality of the writing and on a person reading the record it leaves, the same test Alberta drew a 600 metre line in Bill 42. The bill still has to pass faces in a legislature.

What should a Canadian firm do with a check that can vary?

Write the first 3 policies from rules your firm already holds. A spending cap. A ban on sending client records to personal addresses. A human decision on hiring outcomes, which the job description in Give your AI agent an 8-line job description before access already asks a firm to name. Run the check in dry run mode and read the log each week. A denial your compliance lead cannot explain is a policy that needs plainer words.

Recap: Google’s new check judges each proposed agent action against your written rules and the user’s request, before the action runs, and it logs every verdict. The judge can vary, text replies and streamed responses sit outside it, and the dry run record is the evidence to read first. Put your first 3 policies in writing this week, and read the dry run log before any policy moves to enforcement. The log record starts on the day the engine connects, and no firm can move that start date back.

Total
0
Shares
Prev
Claude can use your team’s computers now. The owner switch comes first
A person working at a desktop computer in an office.

Claude can use your team’s computers now. The owner switch comes first

The setting lands in an admin console this week, and the first person to find it

Next
Rebuilding a public workforce with AI. Copy the training and the guardrails, not the headcount politics
Public service workers training at computers in an office

Rebuilding a public workforce with AI. Copy the training and the guardrails, not the headcount politics

A public service that buys AI before it trains its own people has bought the

You May Also Like