Google is reallocating resources from its Project Mariner browser agent team, signaling a broader industry pivot away from web-based automation toward more powerful, code-driven AI systems. The move comes as competitors like OpenAI and Anthropic gain traction with agents capable of directly manipulating computer systems, rather than relying on simulated human interactions within web browsers.
The Rise of Command-Line Agents
For months, Google Labs staff working on Project Mariner have been reassigned to higher-priority projects, including the development of Gemini Agent. The change reflects a shift in Silicon Valley’s understanding of what constitutes a practical AI assistant. Tools like OpenClaw, which operate through command-line interfaces, are now seen as more efficient and reliable than browser agents that simulate human clicks and scrolling. Nvidia CEO Jensen Huang recently described OpenClaw as a potential “new operating system” for agentic computing.
Browser Agents Struggle to Gain Traction
Early enthusiasm for browser agents – tools like Perplexity’s Comet and OpenAI’s ChatGPT Agent – has waned. As of late 2025, Comet had just 2.8 million weekly active users, while ChatGPT Agent reportedly fell below 1 million. These numbers pale in comparison to ChatGPT’s overall user base, indicating that browser-based automation has not yet resonated with mainstream audiences.
Why the Shift? Computational Efficiency
Experts cite computational limitations as a key factor in the decline of browser agents. These systems rely on processing visual data (screenshots) to understand web pages, which is slow and prone to errors. In contrast, command-line agents work with text-based interfaces, aligning better with the strengths of large language models (LLMs). According to Kian Katanforoosh, CEO of Workera, command-line agents are “10 to 100X less steps to get to the same outcomes.”
New Approaches: Video and Hybrid Systems
Some companies, like Standard Intelligence, are attempting to overcome these limitations by training models on video data instead of screenshots. They claim 50X efficiency gains, even demonstrating a system capable of briefly driving a car autonomously. However, even proponents acknowledge that graphical user interfaces (GUIs) remain essential for tasks that lack programmatic interfaces, like navigating legacy software or healthcare websites.
The Future: Coding Agents Take the Lead
The AI industry is now betting heavily on coding agents – systems that can write and execute code to automate tasks. OpenAI’s Codex and Anthropic’s Claude Cowork are examples of this trend. These agents can manipulate files, create custom software, and integrate with other applications, making them more versatile than browser-based tools. For instance, a coding agent could analyze bank statements and create a personalized financial dashboard.
Despite these advances, mass adoption remains uncertain. Concerns about accuracy and reliability may prevent consumers from automating sensitive tasks like grocery shopping or booking reservations. Nevertheless, the industry consensus is clear: the future of AI agents lies in code, not clicks.
