Google has officially restructured its Project Mariner development team, integrating its experimental browser-automation technology into a unified agent strategy to counter the rise of high-performance systems like OpenClaw. A company spokesperson confirmed that while Project Mariner as a standalone entity is evolving, its “computer use” capabilities are being absorbed into broader products, including the recently deployed Gemini Agent. This pivot signals a tactical shift in Mountain View as the industry moves away from simple browser-based task execution toward deeper operating system integration.
The OpenClaw Effect: A New Operating System for AI
The reorganization follows a massive surge in momentum for “agentic” computing, spearheaded by the open-source tool OpenClaw. Silicon Valley heavyweights now view these tools not merely as plugins, but as the foundational layer for future enterprise and consumer productivity. During a recent developer conference, Nvidia CEO Jensen Huang underscored the magnitude of this shift, comparing modern AI agents to a revolutionary operating system. “Every company in the world today needs to have an OpenClaw strategy,” Huang asserted, highlighting the urgency for tech giants to adapt or risk obsolescence.
The Failure of First-Generation Browser Agents
Despite the initial hype surrounding browser agents at last year’s Google I/O, consumer adoption has remained remarkably stagnant. These tools—designed to click, scroll, and fill forms like human users—have struggled to find a consistent audience. The data reveals a stark reality:
- Perplexity’s Comet: Reached only 2.8 million weekly active users by December 2025.
- OpenAI’s ChatGPT Agent: Reportedly plummeted to fewer than 1 million weekly active users recently.
When compared to the hundreds of millions of users engaging with standard LLM interfaces, browser agent usage currently represents little more than a rounding error in the broader AI ecosystem.
Efficiency Gains: Why the Terminal is Beating the GUI
The industry is rapidly pivoting toward agents like Claude Code and OpenClaw, which interact with computers through command-line interfaces (CLI) rather than graphical user interfaces (GUI). Kian Katanforoosh, CEO of Workera and a Stanford AI lecturer, explains that the visual approach—where an AI takes constant screenshots to “see” a webpage—is computationally expensive and prone to error.
“What Claude Code and OpenClaw showed was that it’s actually much more efficient to work with the terminal,” Katanforoosh noted. He estimates that text-based terminal interactions are 10 to 100 times more efficient than GUI-based actions, as LLMs inherently process text more reliably than visual data.
The 80/20 Rule of Computer Use
While terminal-based agents are winning on speed, some experts argue that GUI interaction remains a technical necessity. Ang Li, CEO of Simular and a former Google DeepMind researcher, suggests an “80/20 split” in the future of agent capabilities. While the terminal handles the majority of tasks, agents must still navigate legacy software and healthcare portals that lack modern APIs. “There will always be problems you have to solve in the GUI,” Li argued, emphasizing that computer use agents fill a critical gap that coding agents alone cannot bridge.
Next-Gen Innovation: From Screenshots to Video Training
The race to optimize computer use has led to breakthrough training methodologies. Startup Standard Intelligence recently unveiled a model trained on video data rather than static screenshots. By utilizing a proprietary video encoder, the company claims its model is 50 times more efficient than previous iterations. To demonstrate this capability, the startup successfully integrated the AI with a vehicle and a keyboard, allowing the model to autonomously navigate San Francisco streets briefly, proving that “computer use” logic can extend into physical world applications.
The Rise of Coding Agents as General Assistants
Major AI labs are now betting that “coding agents” are the true path to general-purpose assistants. By giving an agent the ability to modify files, use various applications, and write bespoke software on the fly, they become significantly more versatile. For instance, a coding agent can transform a raw bank statement into a custom financial dashboard—a task far beyond the reach of a simple browser-clicker.
In response, the competitive landscape is shifting rapidly:
- Anthropic: Launched Claude Cowork, a user-friendly version of Claude Code that removes the need for terminal commands.
- Perplexity: Pivoted from browser-centric tools to a new Personal Computer product.
- OpenAI: Executives are reportedly working to integrate Codex more deeply into ChatGPT to power internal general-purpose agents.
While these advanced capabilities target developers today, Google and OpenAI maintain that the endgame is consumer convenience—automating everything from grocery orders to complex travel bookings. However, the industry remains cautious; until these agents achieve near-perfect reliability, widespread consumer trust remains the final hurdle.
