Gemini Spark: Google's 24/7 AI Agent — I/O 2026 Developer Guide

At 10am PT today, Sundar Pichai unveiled the product every major AI lab has been racing to ship: a persistent, 24/7 personal AI agent that handles long-horizon tasks in the background, integrates with Gmail and Google Docs at a depth no API call can replicate, and connects to third-party services via MCP. Google calls it Gemini Spark — and the technical details buried in today’s I/O keynote are more significant for developers building agentic systems than the last three Gemini model releases combined.

Spark is not a new model. It is an agent: a persistent runtime that runs on dedicated Google Cloud virtual machines, stays online continuously, and takes autonomous action across your digital workspace. Google built it on Gemini 3.5 Flash — also announced today — and a previously internal platform called Google Antigravity, version 2.0 of which is available to external developers starting now.

The product rolls out to Google AI Ultra subscribers in the US starting next week. This guide covers what developers need from today’s announcement: the Spark architecture, the Antigravity 2.0 developer platform, what Gemini 3.5 Flash changes at the model layer, how MCP integration works, and where the opportunity is for developers building agent applications right now.

Spark vs. Gemini Assistant: What Actually Changed

“Agent” is overloaded. Every AI assistant claims to be agentic. Gemini Spark is worth examining precisely because its architecture differs meaningfully from what Google has shipped before.

The original Gemini assistant was stateless: send a query, receive a response, session ends. Extensions and function calling added the ability to invoke tools during a conversation, but each interaction remained bounded by a request lifecycle. Spark breaks that model. It runs persistently on dedicated Cloud VMs. It maintains goal state across hours and days. It can monitor your inbox on a schedule, cross-reference incoming emails against deadlines you described weeks ago, and draft responses before you’ve opened Gmail that morning. The session never ends.

Google describes this as “long horizon” execution — the ability to maintain context and goal state across time windows that no synchronous API call can span. A standard Gemini API call has a lifecycle measured in seconds. Spark’s lifecycle is measured in hours and days. That gap is the entire product.

At launch, Spark integrates with Gmail, Google Docs, Slides, and Sheets. Third-party integrations with Canva, OpenTable, and Instacart ship on day one. MCP-based expansion — covering GitHub, Notion, Slack, and any service with an MCP server — arrives over the summer. But for developers, the more significant announcement is what powers it underneath: Antigravity 2.0.

Google Antigravity 2.0: The Developer Platform Behind Spark

Google Antigravity is the AI-native development platform the company uses internally to build Spark and its other production agent systems. Version 2.0, released today, opens it to external developers as a standalone desktop application with a CLI and SDK.

Three layers define the architecture.

The agent harness wraps Gemini model calls with infrastructure for goal persistence, task decomposition, tool orchestration, safety constraints, and state recovery. This is the component Google describes as preventing agents from “going rogue” — a constraint system that bounds autonomous action to what you explicitly authorize. The harness is available via the Antigravity SDK, meaning you can use the same infrastructure that powers Spark in your own applications.

The orchestration layer is new in version 2.0. The desktop app lets you spawn multiple autonomous agents in parallel, assign them different goals, and have them coordinate via a shared state store. Each agent runs on its own goroutine inside the Antigravity runtime; message-passing between agents is handled by the platform rather than custom coordination code. The CLI exposes the same primitives for terminal-first developers.

The built-in MCP gateway handles server discovery, authentication, connection management, and tool routing for all connected MCP servers. You configure endpoints in a manifest file; Antigravity manages reconnection on failure, request batching, and error handling. Building a multi-tool agent that connects GitHub, Notion, and a custom internal API becomes a configuration task, not an infrastructure one.

# Antigravity 2.0 agent manifest (simplified)
name: "pr-review-agent"
model: gemini-3.5-flash
persistence: cloud     # runs on Google Cloud VMs 24/7

tools:
  mcp_servers:
    - endpoint: "https://github.mcp.io"
      auth: oauth
    - endpoint: "https://notion.mcp.io"
      auth: bearer_token
    - endpoint: "http://localhost:3001"   # custom internal server
      auth: none

goals:
  - name: "daily-pr-digest"
    trigger: cron("0 9 * * 1-5")          # weekdays 9am
    task: >
      Review all PRs opened since yesterday.
      Summarize findings, flag blocking issues,
      and post digest to #eng-reviews Slack channel.

safety:
  max_tool_calls_per_goal: 50
  require_user_confirm: ["git_push", "send_email", "payment_*"]
  allowed_data_sources: ["github.com", "notion.so", "localhost"]

The manifest-driven design is deliberate. Antigravity 2.0 targets the developer who wants production-capable autonomous agents without writing custom orchestration infrastructure. The safety constraints are first-class citizens — defined in the manifest, enforced at runtime. An agent that unexpectedly expands its tool usage hits the configured limit automatically rather than silently overrunning its scope. That behavior is not free in a hand-rolled agent loop; in Antigravity, it is the default.

The same logic that makes the MCP ecosystem compelling applies here: configure the connections, define the goals, let the harness handle coordination and constraint enforcement.

Gemini 3.5 Flash: What Changed at the Model Layer

Spark runs on Gemini 3.5 Flash, announced today alongside the agent product. The headline numbers from Google: 4x faster output token throughput compared to other frontier models in the same tier, with coding, agentic, and multimodal benchmark scores that surpass Gemini 3.1 Pro.

That second claim carries more weight than it appears to. Gemini 3.5 Flash — a Flash-series model — now outperforms the previous Pro-tier model on the benchmarks Google considers most relevant to agentic applications. This extends a pattern that started with Gemini 2.5 Flash: the Flash series has closed the quality gap with Pro while retaining its speed advantage, to the point where Flash is now the correct default for most production workloads rather than a cost-cutting compromise.

Inside the Antigravity runtime specifically, Google reports an optimized variant of Gemini 3.5 Flash running at 12x standard frontier model throughput. The optimization combines speculative execution — the harness predicts likely tool call sequences and pre-fetches responses — with model compression tuned specifically for agentic task patterns. For long-running background tasks where latency per step matters less than throughput over hours, this changes the cost calculus significantly.

// Gemini 3.5 Flash — drop-in replacement for existing Gemini integrations
import { GoogleGenerativeAI } from "@google/generative-ai";

const genAI = new GoogleGenerativeAI(process.env.GEMINI_API_KEY!);

// Same API surface as gemini-3.1-flash — only the model string changes
const model = genAI.getGenerativeModel({
  model: "gemini-3.5-flash",
  generationConfig: {
    temperature: 0.1,     // lower temp for agentic tasks
    maxOutputTokens: 8192,
  },
});

// 4x faster output, surpasses 3.1 Pro on coding + agentic benchmarks
// API pricing publishes this week — ~40-50% cheaper than 3.1 Pro
const result = await model.generateContent({
  contents: [
    { role: "user", parts: [{ text: "Review this PR and flag blocking issues" }] }
  ],
});

Gemini 3.5 Flash rolls out to the Gemini API today. Pricing publishes to the Google AI developer documentation this week; Google indicated approximately one-third to one-half cheaper than previous flagship models. Gemini 3.5 Pro follows in approximately one month for complex reasoning tasks where 3.5 Flash’s quality falls short.

MCP Integration: The Structural Shift in the Ecosystem

Spark’s MCP integration is structurally the most significant announcement from today’s keynote — not because MCP is new, but because of what Google’s adoption confirms.

Until today, connecting Gemini to third-party tools required either the official extensions framework (limited partner list) or custom API integration code per service. A developer building an agent that needed GitHub, Notion, and an internal service simultaneously faced a fragmented integration surface requiring ongoing maintenance across three separate integration patterns.

MCP changes this at the architecture level. An MCP server built once exposes its tools to every MCP-compatible runtime — Claude Code, Cursor 3, Microsoft Agent 365, and now Gemini Spark. The protocol is the integration layer; the client is interchangeable. A GitHub MCP server built for Claude Code works with Gemini Spark without modification. The MCP tool gateway patterns that apply to Claude Code apply directly here.

With OpenAI (ChatGPT connectors), Microsoft (Agent 365), Anthropic (Claude Code, MCP creator), and now Google all supporting MCP, the protocol transitions from “Anthropic’s open standard” to the de facto connectivity layer for the AI industry. An MCP server you build today has a path to simultaneous distribution across every major AI platform.

// MCP server that works unchanged across Spark, Claude Code, Cursor 3, and Agent 365
import { Server } from "@modelcontextprotocol/sdk/server/index.js";
import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
import { ListToolsRequestSchema, CallToolRequestSchema } from "@modelcontextprotocol/sdk/types.js";

const server = new Server(
  { name: "internal-ticket-server", version: "1.0.0" },
  { capabilities: { tools: {} } }
);

server.setRequestHandler(ListToolsRequestSchema, async () => ({
  tools: [{
    name: "get_open_tickets",
    description: "Fetch open support tickets filtered by status",
    inputSchema: {
      type: "object",
      properties: {
        status: { type: "string", enum: ["open", "pending", "escalated"] },
        limit: { type: "number", default: 20 }
      },
      required: ["status"]
    }
  }]
}));

server.setRequestHandler(CallToolRequestSchema, async (request) => {
  if (request.params.name === "get_open_tickets") {
    const tickets = await fetchFromInternalAPI(request.params.arguments);
    return { content: [{ type: "text", text: JSON.stringify(tickets) }] };
  }
  throw new Error("Unknown tool");
});

// Register once — serves Spark, Claude Code, Cursor 3, Agent 365 simultaneously
const transport = new StdioServerTransport();
await server.connect(transport);

Build MCP servers for your internal tooling now. One server, every major AI runtime. This decision compounds quickly: the work you do this week is still paying dividends when the next runtime adopts MCP six months from now.

What Developers Should Do Starting Today

Antigravity 2.0 is available now. Gemini 3.5 Flash is rolling out to the Gemini API today. Spark reaches AI Ultra subscribers next week. Here is what each timeline means for where to put engineering effort.

If you are evaluating agent orchestration platforms: Install Antigravity 2.0 and read through the manifest specification before committing to a custom orchestration architecture. The multi-agent coordination primitives and built-in MCP gateway remove significant infrastructure work from projects that would otherwise require hand-rolled coordination code. Compare it against your current approach before building anything new.

If you have existing Gemini API integrations: Test Gemini 3.5 Flash as a drop-in replacement for 3.1 Flash this week. The throughput improvement and benchmark gains on coding and agentic tasks make it an immediate upgrade for most applications. The API interface is unchanged — a model string swap is the entire migration.

If you are building internal tooling for AI assistants: Prioritize MCP servers over platform-specific integrations. A well-built MCP server gives you simultaneous distribution across Spark, Claude Code, Cursor 3, and Agent 365 for the same development cost as a single-platform integration. The Google ADK for TypeScript also integrates natively with Antigravity 2.0 for TypeScript-first agent development.

If you are planning enterprise deployments: Spark’s MCP expansion and enterprise-grade access controls are scheduled for later this year. Plan the architecture now, build the MCP server infrastructure, and be positioned to integrate when enterprise-grade Spark ships. Starting compliance and security review today means you are not waiting on governance approval when the product is ready.

The Competitive Picture After Today

Gemini Spark joins a set of persistent AI agents that launched in the past six months: Microsoft Agent 365, ChatGPT operator-tier agents, Perplexity Personal Computer, and Claude Code as a developer-facing persistent runtime. Each occupies a different point on the spectrum from developer tool to consumer product.

Google’s structural advantage for Spark is distribution. The Gemini app reaches hundreds of millions of users. Gmail and Workspace integration is native, not bolted on via API. Android OS-level access — announced for Spark on desktop this summer — gives it a file system and local app integration surface that web-based agents cannot match. No competing lab has this combination at this scale.

The open question is enterprise execution. A personal AI agent that touches your Gmail is a product a consumer adopts in a weekend. A Spark deployment touching a corporate inbox requires data residency guarantees, audit logging, access controls, and admin governance that typically take quarters to implement. The labs that win the enterprise agent market will be the ones that make deployment trustworthy, not just the ones that shipped the consumer version first.

Watch Google Cloud Next for enterprise-specific Spark features. If workspace isolation and admin controls ship there, the enterprise agent timeline compresses significantly. If they are absent, competitors have a clear window to close the gap.

The Next 30 Days

Gemini 3.5 Flash pricing details publish this week at the Google AI developer portal. Antigravity 2.0 SDK documentation is available today. Spark reaches Google AI Ultra subscribers in the US next week. Gemini 3.5 Pro arrives in approximately one month. Spark MCP connector expansion and desktop file system access roll out over the summer.

Run the Antigravity CLI on an existing project this week. The MCP gateway configuration takes under an hour if you already have an MCP server running. You will learn more from that session than from any amount of reading the launch announcement. The manifest specification in Antigravity 2.0 reflects production experience with real failures — understanding how Google constrains autonomous action at the harness level is useful independent of whether you adopt Antigravity in your own stack.

Developer tools, templates, and starter kits for building on the Gemini API and MCP-based agent architectures are available at wowhow.cloud — pay once, ship forever.