On May 4, 2026, Google quietly retired Project Mariner and absorbed its browser-automation technology into Gemini Agent, the company’s broader play for integrated personal AI. The move ends an 18-month experiment with a standalone Chrome agent and signals that browser-only automation is no longer viewed as a viable product surface inside Google. Gemini Agent now inherits Mariner’s web-control engine while running where Google believes the next round of agent traction will happen: inside Gmail, Calendar, Docs, Workspace, and Search. The shift narrows Google’s agent strategy at the same moment Anthropic, OpenAI, and Meta are converging on the same conclusion.
What’s actually new
The Mariner website now displays a single notice: the experimental browser agent was discontinued on May 4, 2026, with its technology folded into Gemini Agent. There is no migration tool, no extended sunset window, and no separate roadmap. The team behind Mariner had already been redeployed weeks earlier to work on Google’s response to OpenClaw, the agent framework that went viral in early 2026 after OpenAI hired its creator Peter Steinberg. The reorganization confirms what testers had been guessing since March: Mariner was being kept alive on autopilot while Google rebuilt its agent stack around Gemini 3.1 Ultra.
Gemini Agent is the new front door. It runs as a first-party assistant inside Workspace, Search, and the Gemini app rather than as a Chrome extension. Mariner’s web-action engine — the part that read the DOM, planned multi-step browser flows, and clicked through forms — now powers the “browse and act” tier of Gemini Agent. The integration hooks Mariner used for tab control, cookie handling, and visual grounding remain, but they sit behind a unified planning layer that can also pull from Gmail threads, Calendar slots, Drive files, and Maps locations in the same task.
Three product implications followed inside hours of the announcement. First, Chrome’s agent UI is being absorbed into a Gemini side panel that appears across Google surfaces, not just inside the browser. Second, Workspace administrators get a single agent control plane instead of negotiating Mariner permissions and Workspace permissions separately. Third, developer access to the Mariner-style web automation is moving to the Gemini API rather than a separate Chrome-extension SDK, which simplifies the story for partners but resets timelines for any product that was building against Mariner directly.
Why it matters
- Browser-only agents are losing the platform fight. Mariner, OpenAI’s Atlas, and the early version of Anthropic‘s Computer Use all pitched the browser as the universal substrate. Adoption has been thin. The winning surface in 2026 is the integrated assistant living inside email, calendar, and docs, with browser control as one capability rather than the whole product.
- Google now has one agent story, not three. Before May 4, customers were juggling Mariner (browser), Project Astra (multimodal device), and Gemini Agent (workspace). The consolidation simplifies sales conversations and lets Google ship a single SLA, single audit log, and single billing line for autonomous task completion.
- The Gemini API gets a real agent endpoint. Pulling Mariner’s automation engine inside the API closes a gap with Anthropic’s Computer Use and OpenAI‘s Operator-style API. Developers who wanted to script web tasks against Google’s models had to chain together Vertex AI, Search Grounding, and a third-party browser harness; now there’s a first-party path.
- Workspace becomes Google’s agent moat. The most defensible distribution Google owns is 3 billion Workspace seats. Embedding Gemini Agent there directly — with native access to mail, calendar, drive, and admin policy — is harder for Anthropic and OpenAI to match without a Workspace equivalent.
- Mariner alumni get a chance to ship at scale. The team’s web-action research had real wins: visual grounding on noisy DOMs, robust form handling, recovery from anti-bot challenges. Inside Gemini Agent, that work reaches a much larger user base than a Chrome extension ever could.
- The agentic-commerce timeline moves up. With browser automation and Workspace identity in one product, Google’s path to “book this trip / order these supplies / negotiate this renewal” inside Search and Gmail is shorter. Expect agent-commerce demos to dominate Google I/O later this month.
How to use Gemini Agent today
If you used Mariner in the Chrome extension, the migration is simpler than feared. Three steps get you running with the new browser-control tier inside Gemini Agent.
- Activate Gemini Agent in your Google account. Sign in to
gemini.google.comand turn on the Agent panel. Workspace admins enable it under Apps > Google Workspace > Gemini > Agent features. Personal accounts see it under Gemini Settings. - Grant the right scopes. The browse-and-act tier needs explicit permission for browser automation, plus whichever Workspace data your tasks will touch. Granting at task time is supported but slower; granting up front is cleaner for repeat workflows.
- Hand off your task. Type a goal, not a click sequence. Gemini Agent plans the steps, confirms anything irreversible (purchases, deletions, sends), and reports back with a transcript you can replay or audit.
From the Gemini API, the same capability is exposed as a tool inside the standard generation endpoint:
POST https://generativelanguage.googleapis.com/v1beta/models/gemini-3.1-ultra:generateContent
{
"contents": [{
"role": "user",
"parts": [{ "text": "Find the cheapest direct flight MIA -> SFO next Friday and hold a seat." }]
}],
"tools": [{
"agent": {
"browser": { "enabled": true, "headless": true },
"workspace": { "calendar": "primary", "gmail": "send-on-confirm" }
}
}],
"generationConfig": { "temperature": 0.2 }
}The response stream now includes step events for plan, navigate, click, type, and verify. Each step ships with a screenshot reference so you can render an audit trail without scraping the agent yourself. Builders coming from Mariner’s extension SDK should plan to retire the local Chrome runner; the API path is faster, sandboxed, and survives browser restarts.
How it compares
The browser-action capabilities now span four major vendors. The relevant differences are surface (where the agent runs), distribution (how end users reach it), and pricing model. The table summarizes where each product sits as of early May 2026.
| Product | Surface | Distribution | API access | Pricing model |
|---|---|---|---|---|
| Gemini Agent (Google) | Workspace, Gemini app, Search side panel | 3B Workspace seats + consumer Gemini | Gemini API tool | Per-token + per-step on browse |
| Computer Use (Anthropic) | Claude.ai, API, Cowork | Direct + AWS Bedrock + GCP | Anthropic API tool | Per-token, screenshots metered |
| Operator / Atlas (OpenAI) | ChatGPT, OpenClaw, Atlas browser | ChatGPT users + API | OpenAI Responses API | Per-token + per-action |
| Hatch (Meta, in testing) | WhatsApp, Instagram, Messenger | Internal testing through Q2 | Not yet exposed | Not yet announced |
Two takeaways from the comparison. First, every serious agent platform is now anchored to a productivity or messaging surface, not a browser. Mariner’s death made that explicit but the trend was already visible at Anthropic and OpenAI. Second, the API surfaces are converging on the same shape — a tool block inside a chat completion that exposes browser control plus app integrations — which makes vendor swaps easier than they were nine months ago.
What’s next
Google I/O lands in mid-May 2026 and Gemini Agent is the obvious headline. Expect three demos: an agentic-shopping flow that completes a purchase across Gmail, Maps, and a third-party site; an enterprise agent that triages a Workspace inbox, drafts replies, and books meetings under admin policy; and a developer keynote showing the new agent tool inside the Gemini API with Vertex AI cost controls. The pricing question is the one to watch — Google has not yet published per-step pricing for the browse-and-act tier, and Workspace customers are sensitive to per-seat creep.
The longer-term watch item is whether Google can deliver Workspace-native agent quality before OpenAI’s “super app” or Anthropic’s Cowork close the integration gap. Anthropic shipped Cowork’s Workspace and Microsoft 365 connectors in March. OpenAI’s super app, expected late summer, is rumored to ship with calendar, mail, and file connectors out of the box. Google’s defense is depth of integration plus identity, but it has to materialize quickly. The Mariner consolidation buys the team six months of focused engineering. Whether that’s enough depends on what ships at I/O — and whether Workspace admins trust an agent enough to grant it write access to their orgs at scale.
Frequently Asked Questions
Does Gemini Agent still work as a Chrome extension?
No. The Mariner Chrome extension was deprecated on May 4, 2026, and stopped receiving updates the same day. Browser automation now runs inside the Gemini Agent service, accessible via the Gemini app, Workspace, and the Gemini API. Existing Mariner users were emailed migration instructions; the side panel inside Gemini is the closest one-to-one replacement.
What happens to my Mariner task history?
Google said personal-account task history is migrated automatically into Gemini Agent’s transcript view. Workspace customers with retention policies see history under Vault, controlled by the same retention rules as Gmail and Drive. Anything older than 18 months that was opted into deletion has already been purged per the original Mariner data policy.
Is Gemini Agent available outside the United States?
The browse-and-act tier launched in 35 countries on May 4, with broader EU rollout staged through June pending DSA review. Workspace customers in regulated industries can pin agents to specific data residency regions. The Gemini API exposes the agent tool globally except in jurisdictions where Google has not yet completed local certification.
How does Gemini Agent handle confirmation for sensitive actions?
By default, anything that costs money, sends a message under your identity, or deletes data triggers a confirmation prompt with a side-by-side preview of what the agent intends to do. Workspace admins can tighten this further — for example, requiring step-up authentication on any send-mail action. The agent transcripts log every confirmation event for audit.
Can developers still build a custom agent with Mariner-style web control?
Yes — the path is now the Gemini API agent tool rather than the Mariner extension SDK. The tool exposes browser control as a structured capability, returns step-level events with screenshot references, and supports headless or headful execution. Quotas are higher than Mariner’s experimental limits, and pricing combines per-token model usage with per-step browse charges.
Does this affect Project Astra or Google’s other agent work?
Astra continues as Google’s multimodal device-side assistant for phones and glasses. Astra and Gemini Agent share the same Gemini 3.1 Ultra brain but address different surfaces — Astra is real-time and on-device, Gemini Agent is task-oriented and cloud-resident. The two are expected to integrate further at I/O, with Astra triggering Gemini Agent tasks on the user’s behalf.