Google Reshuffles Browser Agent Team as Industry Shifts Toward Coding and Terminal‑Based AI Agents

Google is reorganizing the team behind Project Mariner, its experimental browser‑automation AI, as the company integrates the technology into broader agent products like Gemini Agent. The move reflects a wider industry pivot toward more efficient terminal‑based agents such as OpenClaw and Claude Code, and toward coding agents that can manipulate software and files. While early browser agents from Google, OpenAI and Perplexity struggled to gain mass adoption, newer models from startups like Standard Intelligence promise higher efficiency. Executives from Google, Nvidia, and AI startups comment on the evolving role of computer‑use agents in consumer applications.

Team Changes at Google

Google is reshaping the group that built Project Mariner, an AI prototype designed to navigate the Chrome browser and perform tasks on a user’s behalf. According to two insiders, several Google Labs staff members who contributed to the research prototype have been reassigned to higher‑priority projects. A Google spokesperson confirmed the reallocation, noting that the computer‑use capabilities developed under Project Mariner will be folded into the company’s overall agent strategy and incorporated into products such as the recently launched Gemini Agent.

Industry Context: Rise of Terminal‑Based Agents

The internal shift comes as AI labs across Silicon Valley respond to the emergence of powerful agents like OpenClaw. Nvidia CEO Jensen Huang likened OpenClaw to a new operating system for “agentic computers,” urging every company to develop an OpenClaw strategy. These agents differ from earlier browser‑automation tools by controlling computers through text‑based terminals, a method that researchers say is far more efficient than processing screenshots of graphical interfaces.

Challenges for Browser Agents

Browser agents initially generated excitement at events such as Google’s I/O conference, where CEO Sundar Pichai highlighted Project Mariner as a potential next‑big bet. Products from OpenAI and Perplexity promised to automate online tasks by clicking, scrolling, and filling out forms. However, user adoption has lagged. Perplexity’s Comet browser agent recorded roughly 2.8 million weekly active users in December 2025, while OpenAI’s ChatGPT Agent fell below one million weekly active users in recent months. Compared with the hundreds of millions of weekly ChatGPT users, browser‑agent usage remains a small fraction of overall AI engagement.

Efficiency Gains from Terminal Interaction

Experts argue that terminal‑based agents require fewer computational steps because both the agents and the underlying language models operate on text. Kian Katanforoosh, CEO of the AI upskilling platform Workera, explained that working with a terminal can be “10 to 100 × less steps” than processing visual screenshots. Startups are also exploring new approaches to improve efficiency. Standard Intelligence released a computer‑use model trained on video streams, claiming a 50 × efficiency boost over prior screenshot‑based models and demonstrating the capability by briefly driving a car autonomously in San Francisco.

The Continuing Role of GUI Agents

Despite the momentum behind terminal agents, some leaders see a lasting need for graphical‑user‑interface (GUI) automation. Ang Li, CEO of Simular and former Google DeepMind researcher, described an “80/20 split” where most problems can be solved via the terminal, but certain legacy applications—such as health‑care insurance portals— still require GUI interaction because they lack accessible APIs.

Shift Toward Coding Agents

Across the AI landscape, firms are increasingly betting on coding agents that can not only operate applications but also modify files and generate custom software. These agents can, for example, ingest bank statements and build a personalized budgeting dashboard. OpenAI plans to power general‑purpose agents inside ChatGPT with its Codex technology, while Anthropic has launched Claude Cowork, an offshoot of Claude Code that does not require a terminal. Perplexity, which previously focused on browser agents, recently introduced a product called Personal Computer that aligns with this coding‑agent trend.

Implications for Consumers

Both Google and OpenAI have suggested that future consumer agents could handle everyday tasks such as ordering groceries from Instacart or reserving a dinner table. While these use cases sound convenient, widespread adoption may hinge on users’ confidence that agents will perform reliably without errors. The industry’s current trajectory indicates a balancing act: maintaining GUI‑based capabilities for legacy tasks while expanding the more efficient terminal and coding‑agent functionalities.