Project Mariner
Last reviewed
May 17, 2026
Sources
28 citations
Review status
Source-backed
Revision
v6 ยท 5,511 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 17, 2026
Sources
28 citations
Review status
Source-backed
Revision
v6 ยท 5,511 words
Add missing citations, update stale details, or suggest a clearer explanation.
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 was 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]
On May 4, 2026, Google shut down Project Mariner as a standalone product, exactly seventeen months after its initial preview. The company folded the agent's underlying technology into Gemini Agent, the auto browse feature in Chrome, and the Gemini 2.5 Computer Use model exposed through the Gemini API and Vertex AI. Google's official Mariner landing page now reads: "Thank you for using Project Mariner. It was shut down on May 4th, 2026 and its technology voyaged to other Google products."[14][15][16]
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]
The name "Mariner" was chosen deliberately by Google to evoke navigation and exploration, and the shutdown notice in May 2026 played on the same metaphor when it said the technology had "voyaged" to other Google products. Both internal and external Google communications repeatedly used nautical language to describe the agent's purpose, treating the open web as the surface to be charted.[14]
Mariner did not appear in isolation. DeepMind had been publishing browser-agent research for several years before the public preview, including work on grounding language models in web environments, agentic reinforcement learning, and visual page understanding. The 2023 paper "A Real-World WebAgent with Planning, Long Context Understanding, and Program Synthesis" and follow-up work on the WebLINX environment fed directly into the system that became Mariner. The agent's observe-plan-act loop is a recognisable descendant of that earlier research line.[2]
Mariner was 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. After May 2025 the system migrated to Gemini 2.5 Pro, and from October 2025 onward production traffic was served by versions of the Gemini 2.5 Computer Use model exposed publicly through the Gemini API.[1][2][17]
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]
Each Mariner session opened a fresh, isolated browser profile in a Google-managed virtual machine. Cookies, logins, and downloaded files for one task were not visible to another, which Google framed both as a safety measure (containment of side effects) and as a productivity feature (the user could run several errands in parallel without polluting their own browsing state). The cloud architecture also unlocked features that the on-device extension could not support, including persistent screenshots, durable task logs that survived a browser close, and the new teach and repeat workflow described below.[3][6]
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]
Teach and repeat was positioned as a halfway point between purely zero-shot agent behaviour and the more brittle record-and-replay automation found in earlier tools such as iMacros or Selenium. Because Gemini interprets the recorded actions semantically rather than as fixed pixel coordinates, the same recipe can in principle survive small visual changes to the underlying website. Google's documentation warned, however, that recipes can still break if a site is redesigned or if it introduces new consent dialogs that were not seen during the original teaching session.[2][10]
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]
On October 7, 2025, Google publicly released the Gemini 2.5 Computer Use model in preview through Google AI Studio and Vertex AI. The model is a specialised variant of Gemini 2.5 Pro that ships through a new computer_use tool in the Gemini API. Google explicitly said the model was the same family of agents that had been powering Project Mariner, the Firebase Testing Agent, and parts of AI Mode in Search, but that this release was the first time the underlying capabilities were made directly available to outside developers.[17][18]
The model is operated inside a developer loop: at each step the application sends the user's request, a screenshot of the current environment, and a short history of recent actions, and the model returns the next action to take. Google reported state-of-the-art results on the Online-Mind2Web, WebVoyager, and AndroidWorld benchmarks at lower latency than rival systems. The Mariner team retired its own bespoke action vocabulary in favour of this shared model, which set the stage for the standalone product's shutdown in May 2026.[17][19]
On November 18, 2025, Google launched Gemini 3 Pro. The new model brought stronger long-horizon planning, better visual understanding of cluttered pages, and direct integration with the agent loop used by Mariner and the public computer use tool. Mariner's cloud worker tier and the experimental Gemini Agent inside the Gemini app were both updated to use Gemini 3 Pro almost immediately.[20]
The same week, DeepMind unveiled Antigravity, an agent-first developer environment that combines an AI editor, an agent manager, and an agent-controlled browser into a single workspace. Antigravity's browser agent shares the visual stack that powered Mariner and lets developers spin up bot-style sessions for end-to-end testing, with the same observe-plan-act loop used by the consumer product. While Antigravity is a separate product targeted at software developers, Google described it as the engineering-grade sibling to Mariner's consumer-grade browsing agent.[21]
Google officially retired Project Mariner on May 4, 2026. The team had been gradually moving the agent's most useful behaviours into mainstream Google surfaces over the preceding six months, and the shutdown notice formalised that migration. The Mariner microsite was rewritten as a single farewell paragraph directing visitors to the Gemini Agent experience, Chrome auto browse, and the public computer use tool in the Gemini API.[14][15][16]
Press coverage treated the shutdown not as a failure but as a reorganisation: the screenshot-driven Mariner stack was judged too slow and too error-prone to compete with code-level and DOM-level alternatives on long, multi-step workflows, but its safety design, its observe-plan-act loop, and its trained policy weights were judged valuable enough to keep alive inside other products. Android Authority called the move a "quiet pivot" rather than a death, and Digital Trends described Mariner as having been "absorbed rather than discontinued."[14][22]
Google's Project Mariner product page broke 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 could manage up to ten parallel tasks since the May 2025 update, with each task running in its own cloud browser. Users could monitor any task from a dashboard, take over manually, or cancel.[3][6]
Google did not publish a formal accuracy figure for end-to-end task completion outside of the WebVoyager benchmark. In hands-on testing, technology reporters described a mix of successes and ordinary failures: the agent could complete linear shopping flows and structured booking forms reliably, but it stumbled on long forms with conditional logic, on aggressive cookie banners that obscured the page, and on sites that detected and blocked automated traffic. Slow round trips between the cloud browser and the model also meant that the agent occasionally timed out before reaching the end of long tasks. The Mariner team acknowledged these weaknesses publicly and cited them as part of the rationale for moving to the lower-latency Gemini 2.5 Computer Use model in late 2025.[5][17][22]
Mariner's action vocabulary was deliberately small. Each step in a task was one of click, scroll, type, navigate, open new tab, switch tab, wait, or done. The model took as input a screenshot of the current viewport, the URL, a short history of recent actions, and the user's overall goal, and returned the next action plus a short natural-language explanation. Compared to systems that ingest a structured DOM, the screenshot-only approach made the agent robust to obfuscated or shadow-DOM sites but added latency because each step required a vision-grade model call. Google later acknowledged this latency penalty as one reason for replacing the screenshot-only pipeline with the hybrid screenshot-plus-element approach used in the public computer use API.[5][17]
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] |
| Gemini 2.5 Computer Use | Google DeepMind | State of the art (Oct 2025) | Google blog, Oct 7, 2025[17] |
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. The October 2025 Gemini 2.5 Computer Use release was the version that finally pushed Google's published numbers above the OpenAI Operator score on WebVoyager, although exact figures vary by harness.[17][19]
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 was focused on browser tasks rather than full operating system control, so it is not always evaluated on OSWorld. With the October 2025 Computer Use release, however, Google did report results on the Online-Mind2Web benchmark, where the new Gemini 2.5 Computer Use model led all reported alternatives, and on AndroidWorld for mobile UI control.[17][19]
Google positioned Mariner as a controlled experiment and placed several deliberate limits on what the agent could 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 said Mariner was 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 was only ever distributed in stages, first to internal testers, then to a small set of external trusted testers, then to AI Ultra subscribers. |
| Sandboxed cloud profile | From May 2025 onward, each task ran in an isolated virtual machine with no access to the user's local files, downloads, or system clipboard. |
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 also said publicly that Mariner's safety guarantees are best-effort. Documentation aimed at testers warned 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 asked testers to use accounts and data they are willing to expose to that risk.[10]
A second category of limits came from outside Google's control. Many large sites use bot-detection services such as Cloudflare Bot Management, Akamai Bot Manager, and PerimeterX to block automated traffic. Mariner sessions identified themselves with a custom user agent and Google asked publishers to allow listed Mariner traffic, but in practice the agent was blocked, rate-limited, or served captchas on a meaningful fraction of pages. Coverage from Human Security, a vendor in the bot-management space, described Mariner sessions as one of several agent-class traffic patterns that site owners began seeing in 2025 and that required new policies to handle.[23]
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]
A partial set of Mariner-derived features, including a slimmed-down Agent Mode and the Gemini in Chrome side panel, later trickled down to subscribers of the cheaper Google AI Pro tier in late 2025. The full ten-task parallel agent capacity remained exclusive to Ultra subscribers until the standalone product was retired in May 2026.[24]
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] That rollout reached general availability on October 7, 2025 with the public preview of the Gemini 2.5 Computer Use model, which is the production stack Mariner itself used during its final months.[17]
Google said it intended to bring Mariner to additional countries beyond the United States, but it never gave a public timeline. The standalone product remained U.S.-only throughout its life. Some Mariner-derived features, including Chrome auto browse and the public computer use API, eventually rolled out more broadly under separate brand names.[3][4][25]
Project Mariner was retired on May 4, 2026. Existing Mariner tasks were drained before the deadline, and AI Ultra subscribers were directed to use the Gemini Agent experience in the Gemini app and Chrome auto browse for equivalent workflows. The Mariner microsite at deepmind.google/models/project-mariner became a farewell page.[14][15][16]
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 at the height of the standalone product's life in 2025.
| Feature | Project Mariner | OpenAI Operator | Anthropic Computer Use |
|---|---|---|---|
| Developer | Google DeepMind | OpenAI | Anthropic |
| Underlying model | Gemini 2.0, later Gemini 2.5 Pro, then Gemini 2.5 Computer Use, finally Gemini 3 Pro | 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 |
| Standalone product retired | May 4, 2026 | Active | Active |
| 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] The October 2025 Gemini 2.5 Computer Use model reset that ranking with state-of-the-art results on Online-Mind2Web and WebVoyager, although those numbers were reported under the new model name rather than the Mariner brand.[17]
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 framed as a deliberate choice to keep the action space and the safety surface small. Mariner is also the only one of the three to have been retired as a standalone product: Operator and Computer Use both remain active brands at the time of writing, while Mariner's capabilities live on inside Gemini Agent and Chrome.[1][5][8][14]
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 said it was engaging with publishers and partners about these questions in parallel with the public rollout.[1][3][5]
Adoption was gated by price and availability. Coverage from outlets such as WebProNews noted that despite the technical progress, the 249.99 US dollars per month AI Ultra subscription and the U.S.-only initial rollout kept Mariner's audience small relative to general consumer chatbots.[13] The eventual shutdown was attributed by several commentators to that adoption gap as well as to the rise of cheaper, code-level alternatives. BigGo Finance and Epinium both framed Mariner's retirement as the end of "the visual web agent era," arguing that screenshot-only agents had been outpaced by DOM-level and tool-level approaches that were faster, cheaper to run, and easier to debug.[26][27]
Not all coverage of the shutdown was negative. Digital Trends and Android Authority framed the move as a normal product consolidation rather than a failure, and Android Headlines noted that the technology behind Mariner had already shipped in higher-volume Google products such as Gemini Agent and Chrome auto browse before the standalone brand was retired. Several commentators argued that Mariner had succeeded precisely because its ideas had been absorbed elsewhere.[14][22][16]
Mariner is part of a wider family of agentic experiments at Google DeepMind:
The Mariner team's most durable contributions are arguably not the consumer product but the building blocks it forced Google to ship. The cloud-hosted browser worker architecture from May 2025 became the basis for Gemini Agent and Chrome auto browse. The teach and repeat workflow informed the macro-recording pattern used in Gemini Agent. The screenshot-driven action model evolved into the public Gemini 2.5 Computer Use API, which is now used by independent companies such as Automation Anywhere, UiPath, Browserbase, Autotab, The Interaction Company, and Cartwheel for their own agentic products.[14][17][18]
The project's safety design also outlived the brand. Confirmation prompts for purchases, the refusal to enter payment details on the user's behalf, and the trained resistance to prompt injection became templates that Google replicated in Gemini Agent and in third-party products built on the computer use API. The Mariner trusted-tester program, which gathered red-team feedback on hostile pages for more than a year before broad rollout, also set a precedent that competitors have since copied.[10][14][17]
Finally, Mariner clarified the practical limits of screenshot-only agents on the modern web. Industry coverage of its retirement noted that the next generation of agents tends to mix visual reasoning with DOM access, with code-level tool calls, and with model-driven planning over structured task graphs. Google's own move from Mariner's pure-vision pipeline to the hybrid stack inside Gemini Agent and Chrome auto browse is widely read as part of that broader industry shift.[22][26][27]