TL;DR
- Team Restructured: Google is dismantling its Project Mariner browser AI agent team and reassigning staff to other initiatives.
- Category Decline: Browser-based AI agents from Google, OpenAI, and Perplexity have all struggled to retain users.
- Coding Tools Win: Terminal and coding agents like OpenClaw and Claude Code are gaining traction as a more efficient alternative.
- Industry Pivot: OpenAI is shifting away from general-purpose browser automation toward narrower, task-specific tools.
At Google I/O last year, CEO Sundar Pichai staked the company’s agentic credibility on Project Mariner, a Chrome browser AI agent designed to navigate the web on a user’s behalf. Google confirmed it is dismantling the team and reassigning staff to higher-priority initiatives, as browser-based AI agents fail to hold onto users across the entire industry. Browser automation is losing ground to coding and terminal-based tools that fit more naturally with how large language models work.
Some Google Labs staffers who worked on Project Mariner have been reassigned, according to two people familiar with the matter. A Google spokesperson confirmed the changes, saying computer use capabilities developed under the project will be incorporated into the company’s broader agent strategy.
Framing a team dissolution as strategic continuity is the corporate language of retreat, not reinvention – and it positions Google to absorb Mariner’s capabilities quietly while shifting public focus to the Gemini Agent.
The Gemini Pivot
Rather than walking away entirely, Google has already folded some Project Mariner capabilities into Gemini Agent, built on insights from the browser agent project and powered by Gemini 3.1‘s advanced reasoning. Expertise developed under Mariner will feed into other products as well, including agentic features across Google’s product lineup.
Pichai highlighted Project Mariner at I/O when browser agents seemed like the industry’s next major bet. By December 2025, Google was still investing in security architecture for Chrome AI agents to address prompt injection vulnerabilities. Less than four months later, the team is being broken up.
Despite Google’s framing of continuity, the speed of the retreat suggests that internal usage data told a story leadership did not want to repeat publicly.
Why Browser Agents Are Struggling
Google is not alone in pulling back. OpenAI launched ChatGPT Agent with four million weekly active paying users, but the product fell below 1 million users within months, a rounding error compared to the hundreds of millions using ChatGPT weekly.
Perplexity’s Comet browser agent reached just 2.8 million weekly active users by December 2025, underscoring how even newer entrants have struggled to build a sustained audience.
User data across three separate products points to a category-level problem, not individual execution failures. Browser agents work by taking screenshots of a webpage, feeding them into an AI model, then acting – a process that can be slow and unreliable, computationally expensive, and error-prone. Every interaction demands a fresh inference cycle, creating latency that makes the tools feel cumbersome next to doing the task manually.
The Coding Agent Surge
Meanwhile, terminal and coding tools are gaining traction, with Cursor’s AI coding platform expanding beyond traditional IDEs and GitHub Copilot CLI adding specialized agents with parallel execution. Anthropic’s highly popular Claude Code CLI has been leading the way and resulted in more consumer friendly products such as the recently launched Claude Cowork.
Nvidia CEO Jensen Huang compared autonomous agent software OpenClaw to a new operating system for agentic computers at the company’s developer conference, declaring “Every company in the world today needs to have an OpenClaw strategy.”
The divergence between browser and terminal tools is structural, not incidental. Terminal environments are text-based – the same medium in which large language models excel – avoiding the overhead of interpreting visual interfaces that browser agents must contend with.
“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.”
Kian Katanforoosh, CEO of Workera and AI lecturer at Stanford University (via Wired)
Anthropic is building out its coding agents to serve as future all-purpose assistants, and OpenAI has just hired the creator of OpenClaw Peter Steinberger.
What Comes Next
With the browser agent category losing momentum, OpenAI is now shifting away from general-purpose browser agents toward narrower tools rather than pursuing broad browser automation. Having earlier broadened access to its Operator AI agent for web-based workflows, OpenAI now appears to be concluding that narrower, task-specific tools win out over general browser automation.
Google’s next major AI event will test whether the Gemini Agent can deliver on the promise that Project Mariner could not, or whether browser-based automation remains a dead end as coding agents reshape the competitive market.