Agentic AI in Gemini: Mariner, Antigravity, Vertex Agents

Google's agent strategy spans consumer (Gemini app, Mariner), developer (Antigravity IDE, Vertex AI Agent Builder), and platform (Gemini API + function calling). This post is the practical lay of the land. Read the overview first for the conceptual foundation.

The Google agent stack

• Project Mariner — Google's browser-controlling agent. Available in Gemini Advanced / AI Pro / Ultra tiers.
• Antigravity — Google's agentic IDE with Gemini at the center. Multi-file code edits, command execution, browser automation in-IDE. See our Antigravity post.
• Vertex AI Agent Builder — the cloud platform for building production agents with state management, tool registry, evaluation. Enterprise-grade.
• Gemini API + function calling — the developer primitive for programmatic agents.
• Gemini Live / multimodal voice agents — real-time voice agents with vision and screen-share.

Project Mariner: the browser agent

Mariner is Google's answer to ChatGPT Operator: a browser-controlling agent that opens tabs, navigates, fills forms, and reports results. Available in Gemini Advanced subscriptions.

Mariner's strengths:

• Tight Chrome integration — it's literally running Chrome under the hood, so sites it visits behave normally.
• Vision-strong — Gemini's native multimodal nature means screenshot reasoning is excellent.
• Long-running tasks — sustained browser sessions over many steps.
• Google service integration — can pivot from web to Gmail, Drive, Calendar without context switching.

Tradeoffs: Mariner is consumer-product positioned. For programmatic browser agents you'd typically use Playwright + Vertex Agent Builder, not Mariner directly.

Antigravity: the IDE agent

For development work, Antigravity is Google's Claude Code analog. A desktop IDE (VS Code fork) with Gemini-powered agent at the center. Multi-file edits, command execution, browser tools, Workspace integration.

For an honest comparison vs Claude Code, see our Antigravity post. Summary: solid product, especially strong in Google-ecosystem workflows; for iOS-specific dev work Claude Code is still the leader.

Vertex AI Agent Builder

The production-grade Google Cloud service for building, deploying, and managing agents. Handles: agent definition, tool registration, conversation state, deployment, evaluation, monitoring, multi-region availability.

Use Vertex Agent Builder when:

• You're building an agent that needs enterprise SLAs.
• You're already on GCP and want unified billing.
• You need compliance posture Google provides (HIPAA, certain regional regimes).
• You need to compare multiple model versions via evaluation tools.

Gemini API + function calling

The developer primitive. Provide function schemas; Gemini calls them in response to user requests; you execute them; loop.

import { GoogleGenerativeAI } from '@google/generative-ai';
const genAI = new GoogleGenerativeAI(process.env.GOOGLE_API_KEY);

const tools = [{
  functionDeclarations: [
    { name: 'read_file', parameters: { type: 'OBJECT', properties: { path: { type: 'STRING' } }, required: ['path'] } }
  ]
}];

const model = genAI.getGenerativeModel({ model: 'gemini-2.5-pro', tools });
const chat = model.startChat();

async function runAgent(goal) {
  let response = await chat.sendMessage(goal);
  while (true) {
    const calls = response.functionCalls();
    if (!calls || calls.length === 0) return response.text();
    const results = [];
    for (const c of calls) {
      const r = await executeTool(c.name, c.args);
      results.push({ functionResponse: { name: c.name, response: r } });
    }
    response = await chat.sendMessage(results);
  }
}

Pattern is similar to Anthropic / OpenAI — agents are agents. The differences are in tool ecosystem, model behavior, and platform integration.

Where Gemini agents win

• Long context. Gemini's 1M-2M token context is unmatched. Agents reasoning over very large codebases or document corpora benefit.
• Multimodal-native. Vision and video reasoning are first-class, not bolted on. Agents that work with screenshots, video, design files are stronger here.
• Google Workspace. Agents that span Gmail, Calendar, Drive, Sheets are most natural here.
• Cost at scale. Gemini Flash is one of the cheapest capable models — agents that loop many times benefit.
• Vertex platform. Enterprise-grade infrastructure with strong eval tooling.

Where Gemini agents trail

• Tool ecosystem. Smaller than MCP (Anthropic) or the OpenAI plugin ecosystem.
• Developer-IDE story. Antigravity is solid but newer than Claude Code; less polish, smaller community.
• iOS workflow. Weaker than Claude's for native iOS development.
• Vendor stability concerns. Google's product-deprecation reputation continues to weigh on developer confidence.

Best use cases for Gemini agents

• Workspace-spanning workflows. "Read this thread of emails, summarize the asks, schedule follow-up meetings."
• Long-document reasoning. Legal review, large codebase navigation, scientific paper synthesis.
• Multimodal tasks. Video analysis, design QA, accessibility audits of screenshots.
• Enterprise GCP-native agents. Bound into your existing GCP stack.
• High-volume, cost-sensitive agents. Using Gemini Flash for cheap inference.

Getting started with Gemini agents

1. Try Mariner if you have Gemini Advanced — assign a small web task.
2. Try Antigravity on a non-critical project. See our Antigravity post.
3. For programmatic agents: get a Google AI Studio API key, set up the Gemini SDK, build a 2-tool minimal agent.
4. For production agents: use Vertex AI Agent Builder. Manage agent definitions in the console, deploy via API.
5. Compare vs Claude / OpenAI for your specific task. Don't assume one platform is right for everything.

See also: Agentic AI Overview, Google Antigravity, Google Gemini, GCP Deep Dive.