Google Revamps Browser Agent Team as OpenClaw Craze Sparks Industry Shake‑up

Google’s internal re‑allocation of Project Mariner staff reflects a broader pivot toward text‑based agentic computing, a shift that has accelerated since the emergence of OpenClaw. According to Wired, two insiders said that several Google Labs engineers who built the browser‑navigation prototype have been reassigned to “higher‑priority projects” as the company folds Mariner’s capabilities into its Gemini Agent platform. The move signals that Google sees the terminal‑driven model, championed by OpenClaw and Claude Code, as a more scalable path for general‑purpose assistants than the screenshot‑heavy, web‑centric approach that Mariner pioneered.

OpenClaw’s rapid adoption has reshaped industry expectations for AI agents. At Nvidia’s recent developer conference, CEO Jensen Huang likened the tool to “a new operating system for agentic computers,” urging every enterprise to develop an OpenClaw strategy (Wired). The contrast with browser agents is stark: while OpenAI’s ChatGPT Agent and Perplexity’s Comet browser assistant have seen weekly active users dip below one million and 2.8 million respectively (Wired), OpenClaw’s command‑line interface delivers tasks with far fewer processing steps. Kian Katanforoosh, CEO of upskilling platform Workera and a Stanford AI lecturer, explained that “the terminal is text‑based and LLMs are text‑based,” making OpenClaw 10‑to‑100× more efficient than screenshot‑based browsers (Wired). This efficiency gain translates into lower latency and reduced compute costs, addressing a key barrier that has limited browser‑agent adoption.

The computational bottleneck of web‑agent pipelines is well documented. Traditional agents capture a series of screenshots, feed the pixel data into a vision model, and then generate action tokens—a process that is both slow and error‑prone (Wired). By contrast, terminal agents operate on raw text streams, eliminating the need for costly visual encoding. Standard Intelligence’s recent release of a video‑based computer‑use model illustrates the industry’s effort to close the efficiency gap: the startup claims its video encoder compresses visual input into the model’s context window at a 50× improvement over prior screenshot‑based methods (Wired). Even with such advances, the consensus among experts cited by Wired is that text‑only pipelines remain the most reliable for large‑scale deployment.

Google’s strategic response is to embed Mariner’s browser‑automation primitives into Gemini Agent, which already supports a broader suite of agentic functions, including code generation and multimodal reasoning (Wired). By consolidating these capabilities, Google aims to preserve the value of its research while aligning with the industry’s shift toward unified, text‑first agents. Sundar Pichai highlighted Mariner at last year’s I/O, positioning browser agents as the next frontier of AI‑driven productivity (Wired). However, the modest user engagement figures for competing browser agents suggest that the market’s appetite may be limited unless the underlying efficiency problem is solved.

The broader AI landscape is now coalescing around agents that can manipulate operating systems directly, rather than merely interacting with web pages. OpenClaw’s creator, who recently joined OpenAI, has helped catalyze this transition, prompting major players—including Google, Nvidia, and emerging startups—to re‑evaluate their roadmaps (Wired). As the sector moves toward terminal‑centric agents, the relevance of browser‑only tools is likely to diminish, relegating them to niche use cases or hybrid models that combine visual and textual inputs only when necessary.