On May 4, Google quietly shut down Project Mariner, the experimental AI agent that headlined last year’s I/O conference. Its landing page now reads: “Technology voyaged to other Google products.”
The shutdown lands two weeks before Google I/O 2026 (May 19), so most coverage reads it as either a product failure or Google falling behind in the AI agent race. Neither is wrong, but if you line up the right comparison points, a different picture emerges.
Project Mariner works by taking browser screenshots, feeding them to a vision model to identify buttons and text fields, then simulating clicks and keystrokes. It shares the same underlying technology as Anthropic’s Computer Use and OpenAI’s CUA (Computer-Using Agent) — screenshot → vision model → action planning → execute GUI operations. But this technology surfaces in two distinct product forms:
Browser agent (Mariner, Claude for Chrome, auto-browse) runs inside a browser sandbox and can only see web page content. Computer use agent (Anthropic Computer Use, OpenAI Codex Computer Use) needs access to the full desktop — screenshots and mouse/keyboard control — though their deployment methods differ: Anthropic’s Computer Use typically runs in a VM or sandbox, while Codex Computer Use is a macOS desktop plugin that runs directly on the user’s real Mac, requiring Screen Recording and Accessibility permissions.
These two categories face different physical environments, and different fates. The CUA model achieved 38.1% on OSWorld (full desktop benchmark) and 58.1% on WebArena (web task benchmark) — but benchmark scores and product form are two separate things. Know which you’re comparing before you compare.
OpenAI’s Operator launched in January 2025 and shut down in August 2025. One frequently cited reason is unreliable purchase flows. But there’s a more fundamental detail: Operator couldn’t even access ChatGPT.com — its own company’s website was on the anti-bot list.
The deployment model of an independent browser agent inherently turns it into a bot. Whether you use headless Chrome, a cloud-hosted browser, or an isolated VM, you’re creating a brand-new browser session — no cookies, no browsing history, no human behavioral fingerprints. Website anti-bot systems have spent over a decade optimizing to detect exactly this kind of session. Amazon’s preliminary injunction against Perplexity Comet in March is just the legal outlet for this category of conflict — the judge ruled that Comet accessing Amazon accounts had “user permission but not Amazon authorization,” citing the Computer Fraud and Abuse Act. Perplexity has appealed; the Ninth Circuit hearing is May 15.
This creates a structural disadvantage for independent browser agent companies: your engineering resources go not into improving model capability or product experience, but into fighting an arms race against an opponent you don’t know. That opponent — Instagram, Amazon, e-commerce and social platforms — has decades of anti-bot experience, direct economic incentives (protect data, protect tiered pricing, protect ad models), and every evasion strategy you deploy triggers their next detection upgrade.
It’s a headwind category. Not because the demand isn’t real — the demand exists elsewhere — but because the product form itself forces you to do two things at once: build a good agent and fight a war you’re unlikely to win.
Let’s look at demand. The typical use case targeted by independent browser agents — book flights, compare prices, shop — demands extremely high reliability. Users won’t tolerate an occasional wrong purchase, and the difference between “wrong purchase” and doing it yourself is just a few minutes. The cognitive trade-off is terrible. OpenAI shut down Operator, then tried Instant Checkout (let users buy inside ChatGPT instead of sending an agent to websites). That also shut down, but the direction was right — it was trying to bypass browser-based operation entirely.
The real demand is on the other side. Bloomberg Terminal’s API access requires a six-figure extra contract. Legacy patient record systems still running in hospitals, insurance claims software written in the 90s, government internal systems — these won’t suddenly sprout APIs just because AI arrived. The problem is, most of them aren’t web applications. Bloomberg Terminal is a Windows desktop program, not an SPA running in Chrome. You’re automating desktop applications, not web pages — so what you need is a computer use agent, not a browser agent.
The demand is real. The category was wrong.
But how do you make computer use work? Same technology — screenshot → vision model → actions — with a different deployment model, and the bottleneck disappears.
If you don’t open a new browser, but instead let the agent operate the user’s existing Chrome — sharing the same cookies, same IP, same browsing behavior patterns — then from the website’s perspective, there’s no way to distinguish user clicks from agent clicks. Google’s Chrome auto-browse, launched January 28, is exactly this path. It’s not a standalone product; it’s a built-in Chrome feature, available to Google AI Pro ($19.99/month) and Ultra ($249.99/month) subscribers. It runs in the user’s real browser, with the user’s real login state. The anti-bot system doesn’t see a new suspicious session — it sees a normal user doing normal operations, just clicking a bit faster sometimes.
Going one step further to computer use — directly controlling the user’s desktop, operating all applications, not just the browser — amplifies this advantage. The user’s desktop carries genuine human operation traces: mouse movement speed, typing rhythm, tab-switching patterns, window-switching intervals. These signals are nearly impossible for bots to fake, and make bot detection nearly impossible to accurately flag an agent. OpenAI’s Codex Computer Use is a concrete instance of this path — it runs on the user’s real Mac, shares the user’s desktop environment and browser login state, and doesn’t trigger the anti-bot problems that appear in headless environments.
Wired reported in March that Google reassigned Project Mariner team members to build an “OpenClaw-like” agent. That same month, Wired documented the industry-wide pivot: Jensen Huang said “every company in the world needs an OpenClaw strategy”; OpenAI execs want Codex to become ChatGPT’s general-purpose agent engine; Anthropic launched Claude Cowork, a Claude Code variant that doesn’t require opening a terminal.
With the categories separated, the three companies’ trajectories can be realigned:
Anthropic: Computer Use has been in beta for 18 months, with 22% on OSWorld (OpenAI’s CUA benchmark page lists Anthropic Computer Use as the previous SOTA; 22% is the figure cited there; humans score 72.4%). Official guidance: “start with low-risk tasks.” The standalone computer use VM path didn’t work. But Claude for Chrome (an extension running in the user’s real browser) is alive and well, adding Quick Mode in March. The real-browser path is working.
OpenAI: The CUA model hits 38.1% on OSWorld and 58.1% on WebArena. The standalone browser agent (Operator) is dead; Instant Checkout is dead. But OpenAI made a move similar to Google’s in another direction: Codex Computer Use is a macOS desktop plugin running directly on the user’s real Mac, requiring Screen Recording and Accessibility permissions to capture screens and simulate input. OpenAI’s own safety guidance states “If Codex uses your browser, it can interact with pages where you’re already signed in” — the agent uses the user’s real browser session, runs the user’s real desktop apps, and doesn’t face the anti-bot detection problems of an isolated VM. Meanwhile, Codex itself is being positioned as ChatGPT’s general-purpose agent engine, pursuing the coding agent + system integration path.
Google: Mariner is dead. Auto-browse is embedded in Chrome (sharing real sessions, no anti-bot triggers). The team is building an OpenClaw-like system-level agent (sharing the real desktop). Both paths avoid the bot detection trap of standalone products.
One more financial backdrop: Alphabet Q1 2026 earnings show Google Cloud revenue up 63% year-over-year to $20 billion, with a $462 billion Cloud backlog. Google makes money from AI by selling infrastructure, not by selling agent products. Anthropic and OpenAI, without a browser or search engine to monetize, must turn agent products themselves into revenue engines. But even so, neither of them stuck with standalone browser agents.
The better story here isn’t “Google fell behind.” It’s that three companies building the same category of product arrived at the same conclusion: standalone browser agents don’t work, but the path to GUI automation is still open — just not via the headless, cloud-isolated route. Google put Mariner’s technology into Chrome (sharing real sessions) and redirected the team toward system-level agents (sharing the real desktop). Both paths avoid the anti-bot system’s detection trap.