Project Mariner

Project Mariner was a research-prototype AI agent from Google DeepMind that browsed the web and took actions inside a user's Chrome browser, such as clicking links, filling forms, scrolling pages, and navigating between sites, on the user's behalf. Built on the Gemini model family, it was announced on December 11, 2024 alongside Gemini 2.0 and scored 83.5% on the WebVoyager web-task benchmark in a single-agent setup, a result Google described as state of the art. Google operated Mariner as a limited research experiment until May 4, 2026, when it was shut down as a standalone product and its technology was folded into Gemini Agent, the Gemini API (including the Gemini 2.5 Computer Use model), and AI Mode in Search, with some coverage also linking it to Chrome's auto browse feature.^{[1][2][14][16][26]}

Google described Mariner at launch as "an early research prototype built with Gemini 2.0 that explores the future of human-agent interaction, starting with your browser."^[1] The system observes the contents of a browser window, plans the steps required to complete a goal, and then carries those steps out, an observe-plan-act loop that Google positioned as the browser-focused counterpart to Project Astra, its multimodal universal-assistant prototype, and to the broader push toward agentic systems that act on the user's behalf rather than only answering questions.^[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]

On May 4, 2026, Google shut down Project Mariner as a standalone product, about seventeen months after its initial preview. There was no formal announcement; the change was first reported by Wired reporter Maxwell Zeff, based on a notice on the project's landing page. Coverage reported that the agent's underlying technology lives on in Gemini Agent, AI Mode in Search, and the Gemini API, whose Gemini 2.5 Computer Use model (also offered through Vertex AI) had been powering Mariner; some outlets also linked it to Chrome's auto browse feature. The Mariner landing page was rewritten to read: "Thank you for using Project Mariner. It was shut down on May 4th, 2026 and its technology voyaged to other Google products."^[15][16][29]

When was Project Mariner announced?

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. Google framed Gemini 2.0 as "our most capable model yet" and "built for this new agentic era," with Mariner as the browser-facing demonstration of that thesis.^[1] 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 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. TechCrunch later listed Mariner alongside OpenAI's Operator, Amazon's Nova Act and Anthropic's Computer Use as competing web-browsing agents.^[3]

The shutdown notice in May 2026 played on the project's nautical name when it said the technology had "voyaged" to other Google products.^[15][26]

Predecessors and DeepMind agent research

Mariner did not appear in isolation. Google DeepMind researchers had published browser-agent research before the public preview, such as the 2023 paper "A Real-World WebAgent with Planning, Long Context Understanding, and Program Synthesis," although Google did not publicly describe how that work related to Mariner.^[33]

How does Project Mariner work?

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. In October 2025 Google said that versions of its Gemini 2.5 Computer Use model, built on Gemini 2.5 Pro, had been powering Mariner.^[1][2][17]

December 2024: Chrome extension

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.^[5] Google's own announcement acknowledged that the prototype was "not always accurate and slow to complete tasks today."^[1]

May 2025: cloud virtual machines

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]

Teach and repeat

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 said that once agents have learned how to do a task, "they can try to replicate the same workflow in the future with minimal input."^[2][4]

Integration with the Gemini app and Search

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]

October 2025: Gemini 2.5 Computer Use model

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, and this release made the capabilities, previously limited to early testers, available to all developers in public preview.^[4][17]

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.^[17][19]

November 2025: Gemini 3 and Antigravity

On November 18, 2025, Google launched Gemini 3 Pro. The same day, Google introduced Gemini Agent in the Gemini app for AI Ultra subscribers in the U.S., describing it as "built on insights from Project Mariner" and powered by Gemini 3's reasoning.^[20]

The same week, Google unveiled Antigravity, an agent-first developer environment that combines a code editor, an agent manager for delegating tasks, and browser integration for testing and simulation. It is a separate product aimed at software developers.^[21]

May 2026: shutdown and migration

Google officially retired Project Mariner on May 4, 2026. Google had been folding Mariner's capabilities into other products, including the Gemini API, Gemini Agent and AI Mode in Search, over the preceding year. The Mariner landing page was replaced with a farewell notice that suggested Gemini Agent for "complex tasks."^[14][15][26]

Press coverage treated the shutdown more as a reorganisation than a failure. Android Authority and Digital Trends said the screenshot-driven approach was resource-intensive and prone to slow performance and errors, while noting that Mariner's technology would live on in other Google products. TechSpot wrote that the shutdown "feels more like a pivot than a surrender," and Digital Trends reported that Mariner's core features would be absorbed into the Gemini API and Gemini Agent rather than discontinued outright.^[14][15][22]

What can Project Mariner do?

Google's Project Mariner product page broke the agent's behavior into three stages: observe, plan, and act.^[2]

Reported and demonstrated tasks include:

• Online shopping. Adding items from a recipe to an Instacart cart by reading a recipe stored in Google Drive and matching the ingredients to products at a chosen store.^[2]
• Job hunting. Searching the Climatebase climate-jobs board for listings that match a resume the user has uploaded.^[2]
• Service bookings. Hiring help on TaskRabbit by extracting details, such as a recent furniture order, from the user's email.^[2]
• Travel and events. Comparing options across multiple sites and, in cooperation with partners, booking event tickets and reservations through services such as Ticketmaster, StubHub, Resy, and Vagaro.^[3][4]
• General research. Visiting a list of pages, gathering information, and returning a summary or filled-out spreadsheet to the user.^[2]

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]

Reported quality and reliability

Google did not publish a formal accuracy figure for end-to-end task completion outside of the WebVoyager benchmark, and its December 2024 announcement said the prototype was "not always accurate and slow to complete tasks today."^[1] TechCrunch found browser-agent prototypes, Mariner included, to be slow and prone to mistakes.^[3] Shutdown coverage said the screenshot-based approach demanded significant compute and could produce slow performance and errors, such as selecting the wrong option on a page.^[15][22]

Underlying technique

Google did not publish Mariner's full action vocabulary. Coverage described the agent as working from screenshots: it took frequent screenshots of the browser, recognised buttons and text, and then clicked or typed for the user, a process that was slower and more compute-intensive than agents that read structured page data.^[5][14][22] The public Gemini 2.5 Computer Use model works in a similar loop, taking the user's request, a screenshot, and a history of recent actions as input and returning UI actions such as clicking or typing.^[17]

How well did Project Mariner perform on benchmarks?

WebVoyager

Project Mariner scored 83.5% on the WebVoyager benchmark in a single-agent setup, a result Google DeepMind described in its December 11, 2024 announcement as "a state-of-the-art result." WebVoyager, introduced in January 2024 by researchers from Zhejiang University, Tencent AI Lab and Westlake University, evaluates an agent's ability to complete real-world web tasks across 15 popular sites, including Amazon, Booking, GitHub, Google Maps, and Wolfram Alpha.^[1][31]

The table below summarizes how Mariner's reported result compares with figures published by competitors for the same benchmark.

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]

OSWorld and Online-Mind2Web

Mariner's launch material did not report an OSWorld score. OSWorld is a separate benchmark for full-computer use, where OpenAI reported 38.1% for the Computer-Using Agent behind Operator, against a previous state of the art of 22.0% from Anthropic's Computer Use.^[30] 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]

Is Project Mariner safe to use?

Google positioned Mariner as a controlled experiment and placed several deliberate limits on what the agent could do.

Sources for these limits include the original Google blog post, follow-up press coverage, and Google's product help pages.^[1][5][10]

Google DeepMind CTO Koray Kavukcuoglu told TechCrunch at launch that the single-tab restriction was a very intentional decision. 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]

Bot detection and the open web

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. HUMAN Security, a vendor in the bot-management space, wrote that there was no public description of a signed identity for Mariner traffic, which presented as a normal Chrome browser, so site owners could not rely on user agent strings, IP ranges, or robots.txt to recognise and govern it.^[23]

Trusted Tester program

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 Google said it was conducting active research on new types of risks and mitigations, including training the model to prioritise user instructions over third-party prompt injection attempts.^[1] TechCrunch described the testers as a small group of preselected users, and Tom's Guide reported that sign-ups were limited to the U.S.^[5][11]

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] At I/O 2025, Google named Ticketmaster, StubHub, Resy, and Vagaro as partners powering some of these agentic flows in AI Mode.^[3]

How much did Project Mariner cost and where was it available?

Initial release (December 2024)

At the December 11, 2024 announcement, Mariner was available only through the trusted-tester program, with no public sign-up.^[1]

Google AI Ultra (May 2025)

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 the highest level of access in the Gemini app, early access to Agent Mode, expanded storage, and early access to other experimental features.^[3][6][7]

Mariner itself remained tied to the Ultra plan until it was retired in May 2026.^[14] Chrome auto browse, a related agentic feature announced in January 2026, was offered to both AI Pro and Ultra subscribers in the U.S.^[25]

Gemini API and Vertex AI

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 public preview on October 7, 2025 with the release of the Gemini 2.5 Computer Use model, versions of which Google said had been powering Mariner.^[17]

Geographic expansion

Google said it intended to bring Mariner to additional countries beyond the United States, but it never gave a public timeline. 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]

Shutdown (May 2026)

Project Mariner was retired on May 4, 2026. The Mariner landing page at deepmind.google/models/project-mariner became a farewell notice that suggested Gemini Agent for "complex tasks"; by late May 2026 the address redirected to the Google DeepMind home page.^[14][15][29]

How does Project Mariner compare to OpenAI Operator and Anthropic Computer Use?

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.

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][30] 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 keeps the action space and the safety surface smaller. Operator was also retired as a standalone product: OpenAI folded it into ChatGPT agent in July 2025 and sunset the Operator research preview site weeks later, while Anthropic's Computer Use remains available to developers. Mariner's capabilities, likewise, live on inside Gemini Agent and other Google products.^{[1][5][8][14][32]}

Reception

Reaction to Mariner from technology press has been generally positive about the ambition and cautious about the execution. In TechCrunch's hands-on demo, Google Labs director Jaclyn Konzelmann called Mariner "a fundamentally new UX paradigm shift," but the reporter 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 also covered the launch, noting that early-access sign-ups were limited to the U.S.^[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. WebProNews reported in April 2026 that Mariner's adoption lagged, and tied the weak uptake of browser agents in general to the inefficiency of screenshot-driven agents compared with terminal and code-based tools.^[13] TechSpot noted that Mariner was available only through the roughly 250 US dollars per month AI Ultra subscription.^[14] The eventual shutdown was attributed by several commentators to that adoption gap as well as to the rise of cheaper, code-level alternatives. Epinium framed Mariner's retirement as the end of "the visual web agent era," and BigGo Finance tied it to an industry shift toward code-capable tools, both 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.^[14][22][16]

Relationship to other Google projects

Mariner is part of a wider family of agentic experiments at Google DeepMind:

• Project Astra. Astra explores a multimodal universal assistant that uses voice, video, and screen sharing in real time. Astra and Mariner share the Gemini 2.0 foundation but address different surfaces. Astra emphasises a continuously available conversational assistant on phones and glasses, while Mariner emphasises asynchronous browser tasks.^[1]
• Stream Realtime and Live API. Gemini's real-time streaming capabilities, used in Gemini Live and the Multimodal Live API, provide the low-latency model interaction layer that some agentic experiences are built on. Mariner uses Gemini's reasoning rather than the live audio path, but both efforts share underlying model and infrastructure work.^[1]
• NotebookLM. NotebookLM is a research-and-writing tool from Google Labs that grounds responses in user-supplied documents. NotebookLM and Mariner are aimed at different user needs, research synthesis versus web action, but both reflect Google's strategy of building specialised agentic surfaces on top of Gemini.
• Jules. Announced alongside Mariner on December 11, 2024, Jules is an experimental coding agent that integrates with GitHub and works on tasks inside a developer's repository. Jules is the coding counterpart to Mariner's web-action role.^[1]
• Agent Mode in the Gemini app. Agent Mode, introduced at I/O 2025, exposes Mariner-style browsing alongside research and Google app integrations in a single interface for AI Ultra subscribers.^[6][7]
• Gemini Agent. Gemini Agent is the consumer agent feature that Google launched in the Gemini app in November 2025, described as "built on insights from Project Mariner," and pointed Mariner users to after the May 2026 shutdown. At launch it was powered by Gemini 3, connected to Google Workspace apps such as Gmail and Calendar, and could organise an inbox or prepare bookings for the user.^[14][20]
• Gemini 2.5 Computer Use. The public computer use model exposed through the Gemini API and Vertex AI is the developer-facing version of the same engine that powered Mariner during its final year. Developers can use it to build browser and mobile UI agents on their own terms.^[17][18]
• Chrome auto browse. Launched in early 2026 for AI Pro and Ultra subscribers, auto browse brings Mariner-style agentic browsing directly into the Chrome browser. Users can ask Chrome to complete a multi-step task and watch the browser do the work in the foreground or background, with confirmation prompts for sensitive actions. Google announced auto browse on January 28, 2026 for AI Pro and Ultra subscribers in the U.S.^[25][28]
• Antigravity. Antigravity is DeepMind's agent-first developer environment, launched in public preview in November 2025. Its browser integration supports testing, research and context gathering for software engineers rather than everyday errands for consumers.^[21]

Legacy

The Mariner team's most durable contributions are arguably not the consumer product but the building blocks it fed into. Google described Gemini Agent as "built on insights from Project Mariner," and versions of the Gemini 2.5 Computer Use model, which Google opened to developers in October 2025, had been powering Mariner.^[17][20] Early testers of that model included Automation Anywhere, UiPath, Autotab, and Poke, with Browserbase hosting a demo environment and running evaluations.^[4][17]

The project's safety design also outlived the brand. Confirmation before purchases, which Mariner required from its first release, reappeared in Gemini Agent, which Google said would seek confirmation before critical actions like making purchases or sending messages, and in the Gemini 2.5 Computer Use model, which can request end-user confirmation for actions such as making a purchase.^[1][17][20]

Finally, Mariner clarified the practical limits of screenshot-only agents on the modern web. Industry coverage of its retirement argued that agents working through APIs, files, code and structured page data had become faster, cheaper and more reliable than agents that visually read web pages.^{[22][26][27][29]}

Continue reading

• Gemini (language model)Gemini is a family of natively multimodal large language models developed by Google DeepMind, first announced on December 6, 2023, that can reason across text, images, audio, video, and code within a single…Mentioned 69 times above
• Gemini (app)The Gemini app is Google's consumer AI assistant, a chat application available on the web at gemini.google.com and as native Android, iOS, macOS and Windows appsMentioned 69 times above
• CComputer useComputer use is the capability of artificial intelligence models to control computers by viewing screens and performing mouse and keyboard actions, much like a human user would.Mentioned 29 times above
• OOpenAI OperatorOpenAI Operator was an AI agent developed by OpenAI that could autonomously perform tasks through web browser interactions.Mentioned 12 times above
• Google DeepMindGoogle DeepMind is an artificial intelligence research and model-development organization at Google.Mentioned 10 times above
• Antigravity (Google)Antigravity is an agentic software development platform from Google, launched in public preview on 18 November 2025 alongside the release of gemini 3 pro.Mentioned 4 times above

More from See also

• Gemini 2.5 Pro
• Gemini 3 Pro
• Gemini 2.5 Computer Use
• Gemini Agent
• Project Astra
• AI Agents
• Browser agents

Topics:AI AgentsGoogle DeepMind

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22Sundar Pichai, Demis Hassabis, and Koray Kavukcuoglu. "Introducing Gemini 2.0: our new AI model for the agentic era." Google blog, December 11, 2024. blog.google/...google-gemini-ai-update-december-2024
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Google DeepMind. "Project Mariner." deepmind.google/...project-mariner
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13Maxwell Zeff. "Google rolls out Project Mariner, its web-browsing AI agent." TechCrunch, May 20, 2025. techcrunch.com/...ariner-its-web-browsing-ai-agent
4. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Sundar Pichai. "Google I/O 2025: From research to reality." Google blog, May 20, 2025. blog.google/...io-2025-keynote
5. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Maxwell Zeff. "Google unveils Project Mariner: AI agents to use the web for you." TechCrunch, December 11, 2024. techcrunch.com/...ai-agents-to-use-the-web-for-you
6. ^1 ^2 ^3 ^4 ^5Abner Li. "Google bringing Mariner browser agent to Gemini app with 'Agent Mode'." 9to5Google, May 20, 2025. 9to5google.com/...project-mariner-gemini-search
7. ^1 ^2 ^3Google. "Gemini app updates at I/O 2025." Google blog, May 20, 2025. blog.google/...gemini-app-updates-io-2025
8. ^OpenAI. "Introducing Operator." OpenAI blog, January 23, 2025. openai.com/...introducing-operator
9. AIMultiple Research. "AI Agents: Operator vs Browser Use vs Project Mariner." research.aimultiple.com/ai-agents
10. ^Google. "How to use Project Mariner." Google Labs Help. support.google.com/...16270604
11. ^1 ^2Tom's Guide. "Google reveals Project Mariner, a new browser agent that can automate your digital life." December 11, 2024. tomsguide.com/...at-can-automate-your-digital-life
12. ^Ron Schmelzer. "Google Gemini's Agent Mode And Project Mariner Show The Future Of AI Agents." Forbes, May 20, 2025. forbes.com/...ariner-shows-the-future-of-ai-agents
13. ^WebProNews. "Google's Project Mariner: Agentic AI Takes the Wheel on Web Tasks, But Adoption Lags." webpronews.com/...l-on-web-tasks-but-adoption-lags
14. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10TechSpot. "Project Mariner is dead, but Google's browser-controlling AI plans are not." May 2026. techspot.com/...-but-google-browser-controlling-ai
15. ^1 ^2 ^3 ^4 ^5 ^6Digital Trends. "Google pulls the plug on Project Mariner, the AI agent that browsed the web like a human." May 2026. digitaltrends.com/...-browsed-the-web-like-a-human
16. ^1 ^2 ^3Android Headlines. "Project Mariner is Over: Google Folds its AI Web Agent into Gemini and Chrome." May 2026. androidheadlines.com/...n-project-mariner-ai-agent
17. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13Google. "Introducing the Gemini 2.5 Computer Use model." Google blog, October 7, 2025. blog.google/...gemini-computer-use-model
18. ^Google. "Gemini 2.5 Computer Use model." Gemini API documentation. ai.google.dev/...-2.5-computer-use-preview-10-2025
19. ^1 ^2 ^3Abner Li. "'Gemini 2.5 Computer Use' has strong web, Android benchmarks." 9to5Google, October 7, 2025. 9to5google.com/...gemini-2-5-computer-use-model
20. ^1 ^2 ^3 ^4Google. "Gemini 3 brings upgraded smarts and new capabilities to the Gemini app." Google blog, November 18, 2025. blog.google/...gemini-3-gemini-app
21. ^1 ^2Geeky Gadgets. "Gemini 3 Pro Support in Antigravity, Google DeepMind's Agent IDE." November 2025. geeky-gadgets.com/google-deepmind-antigravity
22. ^1 ^2 ^3 ^4 ^5Android Authority. "Google quietly kills Project Mariner as the AI agent race shifts gears." May 2026. androidauthority.com/...t-mariner-shutdown-3664323
23. ^Human Security. "What is Google Mariner and why is it on my website?" humansecurity.com/...google-mariner
24. Google. "Google AI Pro and Ultra subscriptions." gemini.google/subscriptions
25. ^1 ^2 ^3Google. "Chrome gets new Gemini 3 features, including auto browse." Google blog, January 28, 2026. blog.google/...gemini-3-auto-browse
26. ^1 ^2 ^3 ^4 ^5BigGo Finance. "Google Shuts Down Project Mariner: Browser AI Agent Killed as Industry Focus Shifts to Code-Capable Tools." May 2026. finance.biggo.com/...ills_Project_Mariner_AI_agent
27. ^1 ^2Epinium. "Google Kills Project Mariner: The Visual Web Agent Era Just Ended." May 2026. epinium.com/...the-visual-web-agent-era-just-ended
28. ^Sarah Perez. "Chrome takes on AI browsers with tighter Gemini integration, agentic features for autonomous tasks." TechCrunch, January 28, 2026. techcrunch.com/...ic-features-for-autonomous-tasks
29. ^1 ^2 ^3gagadget. "Google shut down Project Mariner and folded its tech into Gemini." May 2026. gagadget.com/...er-and-folded-its-tech-into-gemini
30. ^1 ^2 ^3 ^4OpenAI. "Computer-Using Agent." OpenAI blog, January 23, 2025. openai.com/...computer-using-agent
31. ^Hongliang He, Wenlin Yao, Kaixin Ma, et al. "WebVoyager: Building an End-to-End Web Agent with Large Multimodal Models." arXiv:2401.13919, January 2024. arxiv.org/...2401.13919
32. ^OpenAI. "Introducing ChatGPT agent: bridging research and action." OpenAI blog, July 17, 2025. openai.com/...introducing-chatgpt-agent
33. ^Izzeddin Gur, Hiroki Furuta, Austin Huang, et al. "A Real-World WebAgent with Planning, Long Context Understanding, and Program Synthesis." arXiv:2307.12856, July 2023. arxiv.org/...2307.12856

Show all 33 references