Google Restructures Project Mariner Team as AI Agent Momentum Shifts Away From Browser Tools
WIRED has learned that Google is reorganizing the team behind Project Mariner, its AI agent built to navigate the Chrome browser and complete tasks on a user’s behalf. According to two insiders familiar with the changes, multiple Google Labs employees who worked on the research prototype have been reassigned to the company’s higher-priority AI initiatives in recent months.
A Google spokesperson confirmed the internal team changes, but noted that the computer-use AI capabilities developed under Project Mariner will be integrated into Google’s broader long-term AI agent strategy. Some of these capabilities have already been folded into the company’s other existing agent products, including the recently launched Gemini Agent, the spokesperson added.
The restructuring comes as Google and other major AI research teams race to adapt to the rapid rise of high-performance AI agents like OpenClaw. While most of these tools are currently used primarily by developers, Silicon Valley broadly predicts they will soon power general-purpose AI assistants for both consumers and businesses. Nvidia CEO Jensen Huang framed the buzzy technology as a new operating system for agent-powered computers, telling attendees at the company’s developer conference earlier this week: “Every company in the world today needs to have an OpenClaw strategy.”
Google CEO Sundar Pichai first put Project Mariner in the spotlight during last year’s I/O developer conference. At the time, browser-based agents were widely seen as the industry’s next big bet: OpenAI and Perplexity both launched consumer-facing agents that promised to automate everyday online tasks, able to click, scroll, and fill out web forms just like a human user. But consumer adoption of these browser tools has so far fallen far short of industry expectations.
Perplexity’s Comet browser agent only reached 2.8 million weekly active users by December 2025. Meanwhile, OpenAI’s ChatGPT Agent has reportedly dropped to fewer than 1 million weekly active users in recent months. Compared to the hundreds of millions of people who interact with ChatGPT on a weekly basis, usage of standalone browser agents amounts to little more than a rounding error.
New Agents in Town
Momentum across the AI industry has shifted dramatically over the past year toward command-line tools like Claude Code and OpenClaw (whose original creator was hired by OpenAI). Unlike traditional web-browsing agents, these systems control computers via the command line, an approach that has proven far more reliable for completing complex tasks. Many of these new agents even include computer navigation as just one feature among a full suite of agent capabilities, leaving standalone browser agents looking comparatively limited as a standalone product.
Kian Katanforoosh, CEO of AI upskilling platform Workera and an AI lecturer at Stanford University, explains that one core barrier to adoption for traditional screen-based computer agents is their massive computational demand. Most older agents operate by capturing a sequence of webpage screenshots, feeding those images into an AI model, and generating actions based on the visual input. Processing all that visual data is often slow and produces inconsistent results at times.
“What Claude Code and OpenClaw showed was that it’s actually much more efficient to work with the terminal, because the terminal is text-based and LLMs are text-based,” Katanforoosh said. “It’s probably 10 to 100 times fewer steps to get to the same outcomes.”
This is not to say that browser agents are not improving, or that research into general computer-use AI has hit a dead end.
Last month, startup Standard Intelligence released a new computer-use model trained on video footage, rather than static screenshots. The startup says it built a custom video encoder that can compress video to fit within an AI model’s context window, making the new model 50 times more efficient than earlier computer-use AI systems. To demonstrate its technology, the team connected the model to a car, a live road video feed, and a control keyboard, and the model successfully completed a short autonomous test drive around San Francisco.
Ang Li, CEO of computer-use agent startup Simular and a former Google DeepMind researcher, argues that computer-use agents fill a critical gap in AI agent capabilities, and will always be necessary.
“I do see there always being an 80/20 split. You can use the terminal to solve a lot of problems already, but there will always be problems you have to solve in the GUI (graphical user interface),” he said. “For example, if you want to go to a health care insurance website, or some other legacy software, they often don't have an API that a terminal agent can just call up.”
Still, most major AI labs are broadly shifting their core bets away from standalone screen-based agents and toward coding agents. Even for tasks that do not involve writing code, AI teams have found that coding agents’ ability to interact with other applications, edit files, and build custom tools makes them far more useful for end users. For example, if someone needs help building a personal budget, they can upload bank statements to a coding agent and have it build a custom dashboard to track and analyze their spending habits.
OpenAI executives say they want Codex to power general-purpose agents built into ChatGPT. Meanwhile, Anthropic has already launched a consumer-friendly version of this concept: Claude Cowork, a spin-off of Claude Code that does not require users to open a terminal. Perplexity, which bet heavily on browser agents early on, recently launched a similar competing product called Personal Computer.
While coding agents have already found massive traction among developers, it remains unclear whether their expanded capabilities will drive widespread adoption among mainstream consumers. Google and OpenAI have pitched consumer AI agents as tools that can automatically order groceries from Instacart or book dinner reservations. While that level of convenience sounds appealing in theory, most consumers are unlikely to automate these personal tasks until they are fully confident their agent will not make costly mistakes.
