Google is quietly pulling people off Project Mariner, its browser-based AI agent, and redeploying them elsewhere. The timing is not a coincidence—the entire industry just watched Jensen Huang spend a keynote telling 30,000 developers that the future belongs to something very different.Wired first reported the changes, noting that several Google Labs staffers who worked on Mariner have moved to higher-priority projects in recent months. A Google spokesperson confirmed the reshuffle, adding that Mariner’s computer use capabilities will live on inside the company’s broader agent strategy—some already baked into the recently launched Gemini Agent.
Why browser agents never quite took off
Project Mariner had a high-profile launch at Google I/O last year, arriving alongside similar products from OpenAI and Perplexity. The promise was intuitive enough: an AI that browses the web for you, clicks buttons, fills forms, books things. Useful, in theory.In practice, the numbers were underwhelming. Perplexity’s Comet browser agent peaked at 2.8 million weekly active users last December. OpenAI’s ChatGPT Agent has since slipped below 1 million—negligible against the hundreds of millions who use ChatGPT just to chat. The category never found its mass-market moment.The underlying reason is partly structural. Browser agents work by taking rapid screenshots of a webpage, feeding them into a model, and acting on what they see—a loop that is slow, compute-heavy, and prone to errors. Kian Katanforoosh, who runs AI upskilling platform Workera and teaches AI at Stanford, puts the gap in sharp relief: navigating through a terminal, which is text-based just like the models themselves, can be 10 to 100 times more efficient than doing the same work through a browser GUI.
Jensen Huang has a new favourite word: OpenClaw
That efficiency argument is a big part of why OpenClaw has consumed the industry’s attention. The open-source agent platform—created by Austrian developer Peter Steinberger, who joined OpenAI last month—lets anyone spin up autonomous agents from a terminal in a single command. Those agents can read files, call tools, spawn sub-agents, and work through complex multi-step tasks with little hand-holding.At Nvidia’s GTC conference in San Jose this week, Huang called it “the next ChatGPT” in the same breath as Linux and Kubernetes. Nvidia has since launched NemoClaw, an enterprise wrapper that adds a privacy router, network guardrails, and policy controls—essentially making OpenClaw safe enough to run inside a corporate network without IT having a panic attack.The adoption numbers are hard to ignore. OpenClaw became the fastest-growing open-source project in computing history within weeks of launching. In China, Baidu and Tencent are running public installation events; local governments in cities like Wuxi and Shenzhen are offering subsidies worth up to $720,000 for startups building on the platform.
Every major AI lab is now chasing the same thing
Google is not alone in pivoting. Anthropic already ships Claude Cowork—an offshoot of Claude Code built for people who have never opened a terminal. OpenAI wants its Codex to eventually power general-purpose agents inside ChatGPT. Even Perplexity, which staked its agent bet on the browser, recently launched Personal Computer, a terminal-first product.The category is consolidating around a simple insight: agents that work in text-based environments are faster, cheaper, and more reliable than ones wrestling with a browser. For Google, winding down Mariner’s dedicated team is less a retreat and more a recalibration. The browser agent moment has passed—and right now, every company that matters is trying to figure out its OpenClaw strategy instead.
