Google kept its browser agent inside one watched tab

The conversation

On 11 December 2024, Google published "Introducing Gemini 2.0," naming several prototypes built on the new model, including Project Mariner, described there as an "experimental Chrome extension" that "can understand and reason across information in your browser screen, including pixels and web elements like text, code, images and forms" to "complete tasks for you." Google DeepMind's own technologies page for Project Mariner, archived the day of the announcement, calls it "a research prototype exploring the future of human-agent interaction, starting with your browser," and invited people to "join the trusted tester waitlist" rather than offering general access.

What the documents show

Both documents describe the same operating scope: Mariner "can only type, scroll or click in the active tab on your browser," per Google's post, meaning a tab that loses focus is outside what the agent can act on. The DeepMind page adds behavioral detail consistent with that scope, stating the prototype "navigates and interacts with websites on your behalf," "automates repetitive tasks," and "asks for clarification if it doesn't understand an instruction." On the WebVoyager benchmark, Google's post states Mariner "achieved a state-of-the-art result of 83.5%," a figure it immediately qualifies: the same post says the prototype "is not always accurate and slow to complete tasks today, which will improve rapidly over time." Both documents describe December 2024 access as limited to "trusted testers."

The system boundary

Google's post states Mariner "asks users for final confirmation before taking certain sensitive actions, like purchasing something," placing a defined category of action behind a human step rather than automating it end to end. The single-active-tab restriction is itself a designed boundary, not an incidental limit: it keeps the agent's reach to what a user can see and interrupt, rather than letting it operate across background tabs a person is not watching. Google frames its own safety posture as provisional, "conducting active research on new types of risks and mitigations, while keeping humans in the loop" rather than describing that work as complete.

Where it fails

An 83.5 percent benchmark score, by Google's own account, still leaves roughly one in six WebVoyager tasks unresolved, and the same post's admission that the prototype is "slow" and "not always accurate" describes a gap between the demonstrated capability and a general-release consumer product. A trusted-tester population is also, by definition, not a representative sample of how a wider set of websites and instructions might behave once access broadens.

• Does the task being automated stay within a single browser tab kept in view, or would it need a broader scope Mariner does not have?
• What specific actions in this workflow are classified as sensitive enough to require the stated confirmation step?
• Was the benchmark figure measured on tasks similar to the ones actually being delegated, or on a different task set?

Google's own pairing of a state-of-the-art benchmark claim with an explicit "not always accurate" caveat is a useful template for reading any single-number claim about an agent prototype.