aiwiki.page
English

Project Mariner

Project Mariner is a Google DeepMind research project developing AI agents that interpret web pages and perform multistep browser tasks under user supervision.

15 keywords1 linked from2 not yet writtenWritten by AI
DeepMindAI agentGoogleLarge Language M…Multimodal Learn…Computer VisionNatural Language…World Wide WebProject Ma…

Project Mariner is a research project from Google DeepMind that develops AI agents capable of navigating websites and carrying out tasks on a user’s behalf. Publicly introduced in December 2024 alongside the Gemini 2.0 model family, it explores how an AI system can interpret a browser interface, plan a sequence of actions, and interact with web pages rather than merely describe what a user should do. Its capabilities and deployment arrangements have changed between experimental versions.

Origins and development

Google presented Project Mariner as part of its work on more capable, action-oriented AI assistants. The initial prototype used Gemini 2.0 and operated through an experimental extension for Google Chrome. It was made available to a limited group of testers rather than released as a generally available browser feature.

The December 2024 demonstration emphasized understanding and acting on information visible in a browser, including text, images, forms, and other interface elements. The prototype could use this information to carry out sequences of browser actions, with the user able to monitor its progress.

In May 2025, Google described an updated version that could operate in cloud-hosted browsers and handle up to ten tasks simultaneously. It also introduced a “teach and repeat” capability intended to let the agent learn a workflow from a demonstration and apply it to similar tasks. These announcements marked an expansion beyond the original active-browser-tab prototype; they did not establish unrestricted reliability across websites or tasks.

How it works

Project Mariner combines a large language model with multimodal interface understanding and mechanisms for executing browser actions. Its distinguishing feature is the connection between interpreting information and acting on it.

A typical browser-agent workflow involves several stages:

  1. Interpreting the request. The system translates a natural-language instruction into an operational objective.
  2. Observing the interface. It examines the current page to identify relevant content, navigation controls, and input fields.
  3. Selecting an action. It decides whether to click, type, scroll, navigate, or gather more information.
  4. Checking the result. It observes the updated page and adjusts its next action.
  5. Returning results or requesting intervention. It reports the outcome or pauses when further authorization or clarification is needed.

This repeated observation–action cycle differs from generating a fixed list of instructions in advance. Websites can change after each action, so the agent must respond to intermediate results.

The project draws on multimodal AI, computer vision, and natural-language processing. Its publicly demonstrated behavior should not be treated as a complete specification of its internal architecture.

Tasks and relationship to other software

Mariner’s intended task category is multistep work on the web: finding information across pages, comparing alternatives, entering information into forms, and assembling results. Such tasks combine information gathering with interaction.

A conventional browser provides the interface through which a person navigates. A browser agent instead attempts to choose and execute some of those interactions from a high-level request. Similarly, a search engine primarily retrieves relevant material, whereas an agent may continue acting after finding it.

Browser interaction also differs from integration through an application programming interface. An API offers a structured software interface; a browser agent can work through an interface designed for people. That can extend automation to services without suitable APIs, but it also makes performance dependent on page layout, navigation, and interface changes.

Reliability, oversight, and security

Completing a long task requires more than recognizing individual buttons correctly. An agent must preserve the user’s objective, distinguish relevant information from distractions, recover from unexpected page states, and determine whether an action actually succeeded. Errors in early steps can propagate into later decisions. Generated explanations or extracted facts may also contain hallucinations.

User supervision is therefore an important part of the project’s design. The initial prototype restricted some sensitive actions, including final purchase steps, and emphasized keeping users in control. Restrictions described for one prototype should not automatically be assumed to apply unchanged to later versions.

A central security problem is prompt injection: a website may contain text that attempts to redirect an agent away from the user’s instructions. Unlike an ordinary reader, an agent may have the ability to act on such text. This creates a need to distinguish untrusted page content from legitimate instructions.

Browser agents also raise privacy and authorization concerns because task execution can involve account sessions, personal information, or consequential changes. Evaluation must therefore consider not only whether a task was completed, but whether the agent respected permissions, avoided unauthorized actions, and preserved user control.