Google Quietly Downsizes Its Browser Agent Team as the World Goes Mad for OpenClaw

Google Quietly Downsizes Its Browser Agent Team as the World Goes Mad for OpenClaw

Remember when browser-based AI agents were going to change everything? Google certainly does. The tech giant is now reshuffling the team behind Project Mariner, its web-browsing AI agent, as the industry pivots hard towards command-line coding tools like OpenClaw and Claude Code. Sometimes the future arrives, takes one look around, and decides it would rather be somewhere else.

What Happened to Project Mariner?

Project Mariner was a genuine showstopper at Google I/O 2025. The idea was compelling: an AI agent that could navigate the web on your behalf, filling in forms, booking flights, and generally doing the tedious bits of being online. It even posted an impressive 83.5% score on the WebVoyager benchmark. Not too shabby.

But impressive demos and real-world traction are very different beasts. According to reporting from Wired, several Google Labs staffers who worked on Project Mariner have since been reassigned to higher-priority projects. Google has confirmed the changes, though it insists that computer-use capabilities will be folded into its broader agent strategy, including the Gemini Agent. Translation: the tech is not dead, but it has been asked to share its office.

It probably did not help that access to Project Mariner required Google's AI Ultra plan at $249.99 per month. At that price, you would want the agent to do your laundry as well.

The OpenClaw Effect

So what knocked browser agents off their pedestal? In a word: OpenClaw. Originally published as Clawdbot in November 2025 by developer Peter Steinberger, the open-source coding agent exploded onto the scene when it relaunched in January 2026. It racked up 60,000 GitHub stars in just 72 hours. By early March 2026, it had surpassed 250,000 stars, overtaking React as the most-starred non-aggregator software project on the platform.

Let that sink in. A coding agent became more popular on GitHub than React. The library that half the internet is built on.

Steinberger, clearly having caught some attention, announced he was joining OpenAI on 14 February 2026. Nothing says "happy Valentine's Day" quite like a high-profile hiring announcement.

Browser Agents Are Having a Rough Patch

Google is not alone in rethinking its approach. According to Wired's reporting, OpenAI's ChatGPT Agent reportedly saw significant declines in user engagement after launch, though exact figures remain difficult to verify independently. OpenAI has since consolidated its standalone Operator product into ChatGPT Agent and appears to be pivoting towards specialised solutions, including a dedicated shopping agent.

The broader problem is straightforward: getting AI to reliably navigate the messy, ever-changing web is extraordinarily difficult. Cloudflare CEO Matthew Prince has predicted that AI bot traffic will exceed human web traffic by 2027, which rather neatly illustrates the arms race between AI agents trying to browse the web and websites trying to stop them.

Where the Industry Is Heading

The shift towards command-line and coding agents makes a certain amount of sense. Code is structured, predictable, and operates within well-defined rules. Websites, by contrast, are a chaotic jungle of pop-ups, cookie banners, and layouts that change on a whim.

Anthropic is reportedly developing a feature called Cowork to bring its Claude Code agent capabilities to non-programmers. Meta acquired AI agent social platform Moltbook in March 2026. The whole industry seems to agree that the real opportunity lies in agents that work with code rather than agents that wrestle with web browsers.

The Verdict

Browser agents are not finished, but they have certainly been demoted from "the next big thing" to "a feature within a bigger thing." Google folding Project Mariner's capabilities into Gemini Agent is a pragmatic move, even if it lacks the glamour of a standalone product. Meanwhile, OpenClaw's meteoric rise proves that developers want AI tools that meet them where they already work: in the terminal.

The lesson? In AI, the flashiest demo does not always win. Sometimes the quiet, practical tool running in a command line is the one that changes everything.

Read the original article at source.

Share
D
Written by

Daniel Benson

Developer and founder of VelocityCMS. Got tired of waiting for WordPress to load, so built something better. In Rust, obviously. Obsessed with speed, allergic to bloat, and firmly believes PHP had its chance. Based in the UK.