Google Restructures Chrome Team as OpenClaw Gains Popularity
▼ Summary
– Google is restructuring the team behind Project Mariner, its AI browser agent, with staff moving to higher-priority projects, but its capabilities will be integrated into Google’s broader agent strategy.
– The AI industry’s focus has shifted from browser-based agents to more efficient command-line agents like Claude Code and OpenClaw, which control computers via text.
– Browser agents, which automate tasks like clicking and form-filling, have seen low user adoption, with products from Perplexity and OpenAI attracting only a few million weekly users.
– Experts note that browser agents are computationally heavy and less reliable than text-based terminal agents, which require significantly fewer steps to complete tasks.
– Research into computer-use AI continues, with new approaches like video-based models from startups such as Standard Intelligence claiming major efficiency gains.
Google is reorganizing the team responsible for Project Mariner, its experimental AI designed to operate the Chrome browser and perform tasks automatically. This internal shift reflects a broader industry pivot away from standalone browser-based agents toward more versatile systems that interact directly with computer operating systems. While Google confirms the capabilities developed under Mariner will inform its broader agent strategy, the restructuring signals a recalibration as tools like OpenClaw capture significant developer interest and market momentum.
The changes follow a period where browser agents, which automate actions like clicking and form-filling on websites, were heralded as the next major frontier. Google CEO Sundar Pichai showcased Project Mariner last year, and companies like OpenAI and Perplexity launched similar consumer products. However, user adoption has been surprisingly low. Perplexity’s Comet agent had only 2.8 million weekly active users late last year, and OpenAI’s ChatGPT Agent reportedly fell below 1 million recently. These figures are minuscule compared to the hundreds of millions regularly using standard ChatGPT, indicating that browser agents have struggled to meet commercial expectations.
Industry attention has now swung toward a different class of AI tools. Agents like Claude Code and OpenClaw, which operate through computer command lines rather than graphical browsers, are gaining traction for their reliability and efficiency. Nvidia CEO Jensen Huang recently emphasized their importance, calling OpenClaw akin to a new operating system and stating that every company needs a strategy for such technology. The fundamental advantage lies in compatibility: because both the command line and the large language models powering these agents are text-based, they work together seamlessly. Experts note this approach can be ten to one hundred times more efficient than the multi-step process required for visual browser automation.
A significant hurdle for browser-based computer use agents has been their immense computational demand. They typically function by taking repeated screenshots of a screen, analyzing the visual data with an AI model, and then deciding on an action. This process is often slow and can be unreliable. The text-based terminal interaction model avoids these visual processing complexities, offering a more direct and stable method for task completion.
This does not mean research into computer-visual AI has stopped. Innovation continues, with some startups exploring alternative methods. For instance, Standard Intelligence recently demonstrated a model trained on video data rather than static screenshots. The company claims its video compression technique is fifty times more efficient than prior models. In a striking demonstration, they connected this AI to a vehicle’s controls, a live video feed, and a keyboard, resulting in the model briefly driving autonomously on San Francisco streets. Such experiments show the field is evolving, even as the immediate commercial focus shifts.
The restructuring at Google underscores a strategic realignment. The company is integrating Mariner’s learnings into products like the new Gemini Agent while responding to a market that currently values the power and precision of terminal-based agents over the narrower scope of standalone browser automation.
(Source: Wired)
