Google shutters Project Mariner; folded into Gemini

Google quietly shut down Project Mariner on May 4, 2026, removing the experimental browser‑controlling agent as a standalone product and directing users toward Gemini Agent and AI Mode.

The Mariner experiment — unveiled in December 2024 and highlighted at Google I/O 2025 — let an AI take multi‑step actions inside Chrome, from researching purchases to filling forms. Access had been limited to Google’s AI Ultra testers during its preview.

The shutdown was first noticed when Wired reporter Max Zeff and others saw a new message on Project Mariner’s landing page that reads, “Thank you for using Project Mariner. It was shut down on May 4th, 2026 and its technology voyaged to other Google products.” Google did not issue a separate public announcement.

Technically, Mariner was unusual among agents because it “saw” the web the way a human does: the system took frequent screenshots of an active browser tab, ran visual recognition to find buttons and text, and then clicked or typed to complete tasks. That visual, pixel‑level approach let it operate across sites without special integrations.

But the screenshot‑based model carries costs. Visual processing at scale demands significant compute, can be slow, and is more prone to misclicks or context errors than agents that use structured page data or APIs. Continuous access to a user’s visible tab also raised privacy and safety questions for some testers.

Google says it is not discarding Mariner’s work — the core capabilities are being folded into Gemini Agent, AI Mode in search, and the broader Gemini API. That migration means many of Mariner’s automations will be accessible inside Google’s main AI surfaces rather than as a separate labs experiment.

The move fits a wider industry shift away from visual browser agents toward agents that operate at the file, code, or API level — tools that modify documents, call web APIs, or run code rather than repeatedly parsing screenshots. Competitors and third‑party tools such as OpenClaw and coding‑focused agents have accelerated that trend by showing faster, cheaper, and more controllable workflows.

For end users the practical change should be subtle: routines once tested in Mariner are being re‑exposed through Gemini Agent and AI Mode, which are already integrated into Google’s search and productivity flows and don’t require an Ultra‑only subscription in the same narrow way Mariner did. That promises broader access, but also a different interaction model.

Developers and integrators face a different tradeoff. Google has been folding agentic features into the Gemini API and other developer tooling, which favors building actions as structured tasks or API calls that are easier to observe, debug, and secure than pixel‑level automation. That makes the capabilities more reusable inside apps and enterprise workflows.

Timing matters: the shutdown came a few weeks before Google I/O 2026, which starts May 19, and follows reporting that staffers had been reassigned away from the Mariner team earlier this spring. Observers say the company appears to be consolidating experimental agent work into product teams ahead of larger platform announcements.

Product designers will keep wrestling with the same tradeoffs that Mariner exposed: visual browser agents can work with any site but are heavier and harder to secure, while API‑first or code‑level agents are faster and more predictable but require site cooperation or developer work. Expect Google and others to continue hybrid approaches that mix both where needed.

For customers who liked Mariner’s promise of “send an agent and get it done,” the experience will shift behind different interfaces — a Gemini Agent workflow, an AI Mode search result, or a Gemini API integration — rather than a separate browser extension. That’s a sign Google is productizing research into services it can scale.

Project Mariner’s end is best read as a pivot, not a full retreat: Google preserved the technical work and is folding it into platforms where agents can call APIs, manipulate files, or run code alongside user data and enterprise controls. The episode also marks a clearing in the agent market — one code‑ and API‑centric architecture is becoming dominant.