Google Shakes Up Its Browser Agent Team Amid OpenClaw Craze

When Google CEO Sundar Pichai took the stage at I/O 2025 and demonstrated Project Mariner navigating Chrome on a user's behalf, browser agents looked like the next big thing. Less than a year later, Google is quietly restructuring the team behind that demo — and the broader industry has already moved on.

The shift is stark. Google confirmed to WIRED that some Google Labs staffers who worked on the Project Mariner research prototype have moved to higher-priority projects. The computer-use capabilities developed under Mariner will be folded into other agent products, including the recently launched Gemini Agent, a spokesperson said. The research prototype phase is over.

The cause is equally clear: CLI-based coding agents like OpenClaw and Claude Code have eaten browser agents' lunch.

The numbers are brutal. Perplexity's Comet browser agent reached just 2.8 million weekly active users in December 2025. OpenAI's ChatGPT Agent fell below 1 million weekly active users in recent months. For context, ChatGPT itself has hundreds of millions of weekly active users. Browser agents are a rounding error.

Meanwhile, OpenClaw crossed 250,000 GitHub stars in under four months, surpassing React to become the most-starred non-aggregator project in history — a title React held for years — per star-history.com. Jensen Huang called OpenClaw "probably the single most important release of software, probably ever" at the Morgan Stanley TMT Conference on March 5, 2026. At Nvidia's GTC conference on March 16, he put it in historical context: "Every company in the world needs to have an OpenClaw strategy. Just like they needed a Linux strategy, an HTML strategy, a Kubernetes strategy." At the same GTC appearance, Huang also said Claude Code and OpenClaw "sparked the agent inflection point."

The efficiency gap explains the divergence. Kian Katanforoosh, CEO of AI upskilling platform Workera and a Stanford lecturer, put it plainly: browser agents work by taking screenshots of web pages, feeding them to an AI model, then taking actions based on what it sees. Processing all that visual information is slow and often unreliable. "What Claude Code and OpenClaw showed was that it's actually much more efficient to work with the terminal, because the terminal is text-based and LLMs are text-based," he told WIRED. "It's probably 10 to 100X fewer steps to get to the same outcomes."

The technical explanation is straightforward. Terminal and IDE environments output structured text — the native language of large language models. A coding agent can read a file, run a test, see the error, fix it, and iterate in a tight loop. Browser agents have to transcribe every pixel on screen into tokens, losing structural information in the process, then reverse-engineer what actions to take. It's computationally expensive and semantically lossy.

This isn't to say browser agents are dead. Ang Li, CEO of computer-use startup Simular and a former Google DeepMind researcher, argues that GUI-based agents will always be necessary for legacy software without APIs — healthcare insurance portals, outdated enterprise tools, anything that doesn't expose a clean programmatic interface. "There will always be problems you have to solve in the GUI," he said.

And the browser agent approach is evolving. Standard Intelligence recently released a computer-use model trained on video rather than screenshots, claiming 50X efficiency improvement. The startup even rigged one up to drive around San Francisco briefly.

But the strategic momentum has shifted. OpenAI hired OpenClaw's creator Peter Steinberger in February, with Sam Altman calling him "a genius with a lot of amazing ideas about the future of very smart agents." Nvidia announced NemoClaw at GTC — a security-hardened wrapper for OpenClaw targeting enterprise deployment, with sandboxing and policy guardrails. OpenAI wants Codex powering general-purpose agents inside ChatGPT. Anthropic already built Claude Cowork, a non-terminal offshoot of Claude Code.

Even for tasks that have nothing to do with coding, the CLI agent paradigm is proving more capable. Upload bank statements to a coding agent and have it build a custom spending dashboard? That's not a coding task — but a coding agent's ability to write code, execute it, and iterate is exactly what makes it possible.

For Google, the restructuring of Project Mariner is a recognition that being early isn't the same as being right. Browser automation looked like the obvious agent interface until it wasn't. The team's capabilities aren't going away — they're being redistributed into products where they might actually ship.

The browser agent chapter isn't closed. But the CLI agent era is here, and it's moving fast.