CtrlK
Project Mariner is a research prototype developed by Google DeepMind that builds an AI agent capable of browsing the web on a user's behalf. The system observes the contents of a browser window, plans the steps required to complete a goal, and then carries out those steps by clicking links, filling forms, scrolling pages, and navigating between sites. It was first introduced on December 11, 2024 as part of the Gemini 2.0 launch and was described by Google as an early experiment in human-agent interaction inside the browser.[1][2]
Mariner started as an experimental Chrome extension distributed to a small group of trusted testers. In May 2025, at Google I/O, Google moved the agent to cloud-hosted virtual machines, allowed it to run up to ten tasks at once, and made it available to subscribers of the new Google AI Ultra plan in the United States.[3][4] The product is positioned as a sibling to Project Astra, which explores a multimodal universal assistant, and to the broader push toward agentic systems that act on the user's behalf rather than only answering questions.[1]
Google announced Project Mariner in a blog post titled "Introducing Gemini 2.0: our new AI model for the agentic era," published on December 11, 2024 by Sundar Pichai, Demis Hassabis, and Koray Kavukcuoglu. The same announcement included Gemini 2.0 Flash, an updated Project Astra for a multimodal universal assistant, and Jules, an experimental coding agent.[1] Mariner was framed as the browser-focused entry in this group, alongside Astra (general assistant) and Jules (software engineering).
At launch, Mariner was distributed only as an experimental Chrome extension to a small group of trusted testers in the United States. Google described the project as a research prototype rather than a finished product and said it expected to learn from how testers used the agent before broader release.[1] The team explicitly framed Mariner as part of a long-running line of agent research at DeepMind, including work on reinforcement learning, large language model planning, and multimodal grounding.
The announcement came a few weeks after Anthropic introduced Computer Use for Claude on October 22, 2024, and roughly six weeks before OpenAI launched Operator on January 23, 2025. Press coverage at the time treated the three releases as the opening salvo of a new "agent race" among the major AI labs, with each company pushing a different architecture for putting language models in control of a computer.[5]
Mariner is built on top of Gemini 2.0, Google's multimodal model family released in December 2024. The agent uses Gemini's ability to reason over both pixels and structured web content, including text, code, images, and form elements, to understand what is on screen and decide what to do next.[1][2]
In its initial form, Mariner ran as a Chrome browser extension on the user's own machine. The extension captured screenshots of the active tab and sent them to Gemini in the cloud, which returned a sequence of low-level actions such as moving the cursor, clicking, scrolling, or typing. Those actions were then replayed inside the user's Chrome browser. A sidebar in Chrome showed Gemini's running commentary, including the goal it understood, the steps it planned, and the action it was currently taking.[5]
Because the agent was driving the user's own browser, the user could not freely use Chrome for other work while a task was running. TechCrunch's hands-on demo at launch reported delays of about five seconds between cursor movements, which Google attributed to the slow round trip between Chrome and Gemini and to deliberate pacing for safety.[5]
At Google I/O 2025, Google rebuilt Mariner so that each task ran inside its own browser on a cloud virtual machine rather than on the user's own computer. This change brought Mariner in line with OpenAI Operator and Amazon's Nova Act, both of which use cloud-hosted browsers, and it allowed users to delegate work to the agent and continue using their own machine for other things.[3][4] The cloud architecture also let one user kick off several Mariner tasks in parallel: Google said the new version could run up to ten different tasks at the same time.[3]
The May 2025 update introduced a feature Google calls "teach and repeat." Once a user has guided Mariner through a multi-step browser task, the agent stores the learned routine and can attempt to redo it later with less explicit instruction or apply the same pattern to similar tasks. Google's product page describes this as letting users "teach" Mariner once and then have it "repeat" the workflow on demand.[2][6]
In the same announcement, Google said it would bring Mariner into the Gemini app as a feature called Agent Mode, available to Google AI Ultra subscribers on desktop. In Agent Mode, the Gemini chat appears on the left side of the screen while a live browser running on a Google virtual machine appears on the right, so the user can watch the agent work and intervene at any time.[6][7]
Google also said Mariner's underlying capabilities would power parts of AI Mode in Google Search and would be exposed to developers through the Gemini API and Vertex AI, so third-party applications could build their own agents on the same browser-control stack.[3][4]
Google's Project Mariner product page breaks the agent's behavior into three stages: observe, plan, and act.[2]
| Stage | Description |
|---|---|
| Observes | Reads the active web page, including pixels, text, code, images, and form fields, and identifies elements such as buttons, links, search boxes, and dropdowns. |
| Plans | Interprets the user's goal, breaks it into a sequence of steps, and explains the plan back to the user before acting. |
| Acts | Moves the cursor, clicks, scrolls, types into form fields, and navigates between pages while keeping the user informed about the current step. |
Reported and demonstrated tasks include:
Mariner can manage up to ten parallel tasks since the May 2025 update, with each task running in its own cloud browser. Users can monitor any task from a dashboard, take over manually, or cancel.[3][6]
In the December 11, 2024 announcement, Google reported that Project Mariner achieved a score of 83.5% on the WebVoyager benchmark in a single-agent setup. WebVoyager, introduced in early 2024 by researchers from Tsinghua University and others, evaluates an agent's ability to complete real-world web tasks across about fifteen popular sites, including Amazon, Booking, GitHub, Google Maps, and Wikipedia.[1]
The table below summarizes how Mariner's reported result compares with figures published by competitors for the same benchmark.
| Agent | Developer | WebVoyager (reported) | Source |
|---|---|---|---|
| Project Mariner | Google DeepMind | 83.5% (single-agent) | Google blog, Dec 11, 2024[1] |
| Operator (Computer-Using Agent) | OpenAI | 87% | OpenAI announcement, Jan 23, 2025[8] |
| Computer Use | Anthropic | 56% | Reported by third-party comparisons[9] |
Benchmark scores for browser agents are sensitive to the exact set of tasks, time limits, and harness used, so direct comparison between vendors is approximate.
Mariner's launch material did not report an OSWorld score. OSWorld is a separate benchmark for full-computer use, where Operator has reported 38.1% and Anthropic's Computer Use 22.0%.[9] Mariner is currently focused on browser tasks rather than full operating system control, so it is not always evaluated on OSWorld.
Google has positioned Mariner as a controlled experiment and has placed several deliberate limits on what the agent can do.
| Restriction | Description |
|---|---|
| Active tab only (initial release) | In the December 2024 Chrome extension, Mariner could only act in the foreground tab the user was watching. It could not switch tabs or run when Chrome was minimized. |
| Confirmation for sensitive actions | Mariner asks the user to confirm before steps that could have real-world consequences, such as completing a purchase, accepting terms of service, or agreeing to cookie banners. |
| No automatic payments | The agent will not enter credit card or billing details on its own. Users complete checkouts and password entries manually. |
| Prompt-injection defenses | Google has said Mariner is trained to resist instructions hidden in web pages that try to override the user's actual goal, including phishing-style emails and adversarial site content. |
| Trusted-tester rollout | The system has only ever been distributed in stages, first to internal testers, then to a small set of external trusted testers, then to AI Ultra subscribers. |
Sources for these limits include the original Google blog post, follow-up press coverage, and Google's product help pages.[1][5][10]
Google's CTO and other executives stressed at launch that the single-tab restriction was intentional. Because the agent is taking real actions on a user's behalf, the team chose a step-by-step interaction model where the user can see and stop every move. TechCrunch's hands-on test described the agent as visibly slow, with several-second delays between actions, and noted that it would occasionally stop to ask clarifying questions, such as exact quantities while filling a grocery cart.[5]
Google has also said publicly that Mariner's safety guarantees are best-effort. Documentation aimed at testers warns that the system may fail to ask for confirmation in every case, may misinterpret instructions, and may take incorrect actions on a user's accounts. Google asks testers to use accounts and data they are willing to expose to that risk.[10]
From its launch in December 2024 through the broader rollout in May 2025, Mariner was distributed only through a Trusted Tester program. Testers received an experimental Chrome extension and reported issues to Google through internal feedback channels. Google described this group as small and U.S.-based and used it to study real-world failure modes, especially around prompt injection and unintended side effects on third-party sites.[1]
Google also said it had begun parallel conversations with web publishers and platform owners about the implications of agents acting on user behalf, including how to identify agent traffic, how to handle bot-detection rules, and how to respect site terms.[1] These conversations later evolved into the partner integrations announced at I/O 2025, including ticketing and reservations partners.[3][4]
At the December 11, 2024 announcement, Mariner was available only through the trusted-tester program, with no public sign-up.[1]
On May 20, 2025, at Google I/O, Google made Mariner available to subscribers of the new Google AI Ultra plan, which is priced at 249.99 US dollars per month and is initially limited to U.S. users. AI Ultra also bundles other premium products, including Gemini Advanced, expanded storage, and early access to other experimental features.[3][6][7]
Google said at I/O 2025 that Mariner's computer-use capabilities would be exposed to developers through the Gemini API and Vertex AI so that third parties could build agentic applications on the same stack.[3][4] Google's product page lists this as an ongoing rollout.[2]
Google has said it intends to bring Mariner to additional countries beyond the United States, but it has not given a public timeline.[3][4]
Mariner is one of three high-profile browser- or computer-use agents released by major AI labs between late 2024 and early 2025. The table below summarizes the publicly reported design choices of each.
| Feature | Project Mariner | OpenAI Operator | Anthropic Computer Use |
|---|---|---|---|
| Developer | Google DeepMind | OpenAI | Anthropic |
| Underlying model | Gemini 2.0, later Gemini 2.5 | Computer-Using Agent based on GPT-4o | Claude 3.5 Sonnet and later |
| First public preview | December 11, 2024 | January 23, 2025 | October 22, 2024 |
| Execution environment | Chrome extension on user device (Dec 2024); cloud VMs (May 2025) | Cloud-hosted browser on OpenAI servers | Local sandbox or developer-supplied virtual machine |
| Scope | Web browser only | Web browser only | Whole desktop (mouse, keyboard, screenshots) |
| Parallel tasks | Up to 10 (May 2025 onward) | Multiple, served from OpenAI cloud | One per session |
| Pricing for end users | Bundled in Google AI Ultra at 249.99 US dollars per month | Bundled in ChatGPT Pro at 200 US dollars per month | Pay-per-token via Anthropic API |
| Sensitive-action policy | Asks user to confirm purchases, terms, cookies, passwords | "Watch Mode" supervision and pause on sensitive sites | Sandboxing and developer-defined guardrails |
Reported WebVoyager scores favored Operator at 87%, with Mariner at 83.5% and Computer Use at 56%, although the harnesses and prompts vary across vendors.[1][8][9]
Mariner's most distinctive design choice is the focus on the browser. Anthropic's Computer Use can drive the entire desktop, which gives it broader reach but more attack surface. OpenAI's Operator and Google's Mariner both stay inside a browser, which Google has framed as a deliberate choice to keep the action space and the safety surface small.[1][5][8]
Reaction to Mariner from technology press has been generally positive about the ambition and cautious about the execution. TechCrunch's hands-on demo described Mariner as "a fundamentally new UX paradigm shift" but emphasized that the prototype was slow, with five-second delays between actions, and locked the user out of multitasking in their own browser.[5] Tom's Guide and other early reviewers made similar observations.[11]
Following the May 2025 update, coverage focused on the move to cloud virtual machines and on the new Agent Mode in the Gemini app. 9to5Google described Agent Mode as a desktop split-screen experience for AI Ultra subscribers in which Gemini orchestrates research, browsing, and Google app integrations on behalf of the user.[6] Forbes contributor Ron Schmelzer wrote that Agent Mode together with Mariner suggested how Google intended to package agentic capabilities for paying users.[12]
Industry observers also discussed the broader implications of browser agents for the open web. Some publishers worry that agents acting on behalf of users may reduce direct page views and ad impressions, while others see opportunities to expose structured booking and shopping APIs to agent traffic. Google has said it is engaging with publishers and partners about these questions in parallel with the public rollout.[1][3][5]
Adoption has been gated by price and availability. Coverage from outlets such as WebProNews has noted that despite the technical progress, the 249.99 US dollars per month AI Ultra subscription and the U.S.-only initial rollout have kept Mariner's audience small relative to general consumer chatbots.[13]
Mariner is part of a wider family of agentic experiments at Google DeepMind: