Gemini Spark: Google's Always-On Personal Agent

Google unveiled Gemini Spark at I/O 2026 this morning — the first 24/7 ambient personal agent shipped by a Big Three AI lab. Spark runs on dedicated Google Cloud VMs so tasks continue after a laptop closes or a phone locks, and it integrates natively with Gmail, Calendar, Docs, Drive, and three MCP-connected third-party apps at launch.

Every competing personal AI assistant — Apple Intelligence, Microsoft 365 Copilot, ChatGPT Atlas Agent mode, and Claude — is session-bound. The session ends when the window closes. Spark's persistent-VM architecture is not a feature increment; it is a category shift that moves the personal-agent model from on-demand to ambient. Whether that shift becomes the default UX for knowledge workers or stays a power-user surface is the real question the May 25 beta wave will begin to answer.

This guide covers what actually launched today at I/O, how the persistent-VM architecture works, the full first-party and third-party integration surface, the Antigravity harness underneath, the AI Ultra repricing that landed alongside Spark, and a five-way landscape table comparing Spark to every session-bound competitor in Q2 2026. We also flag the blast-radius risks Google's staged rollout is deliberately hedging against.

Sundar Pichai opened the Google I/O 2026 keynote this morning by describing Gemini Spark as “your personal AI agent in Gemini app that helps you navigate your digital life, taking action on your behalf and under your direction.” Josh Woodward, VP of Google Labs and the Gemini App, followed with a product demo showing Spark drafting a status-update email by pulling context from Gmail threads, Docs, Sheets, and Slides simultaneously — “Need to send an email to your boss with a status update? Spark can pull all the facts from your emails, your docs, your sheets, and slides and write the draft for you.”

The launch cadence is deliberately staged. Trusted testers received access this morning. The wider AI Ultra beta opens the week of May 25. Geographic availability at launch is United States only; no timeline for international rollout has been disclosed. The staged rollout — from the announcement at TechCrunch's launch coverage to the blog.google pricing post by VP Shimrit Ben-Yair — is consistent with how Google has staged AI Inbox and the Daily Brief feature in Gmail over the past year.

Importantly, the pre-event leaks pegged Spark as AI Ultra-only. The actual announcement expanded access to AI Pro as well — both tiers gate Spark at launch. That means the $19.99/mo AI Pro plan is now a viable entry point, though usage limits at Pro are a fraction of what Ultra provides.

The architectural distinction that makes Spark a category move rather than a product feature is deceptively simple: Spark tasks run on dedicated virtual machines on Google Cloud, not on the user's device. When a user closes the laptop, the Spark VM keeps running. When the phone locks, the Spark VM keeps running. A regular Gemini chat session terminates when the browser tab closes; a Spark task is a persistent process with its own state, a trigger system, and a notification pathway back to the user.

The practical consequence is that Spark can monitor an inbox in real time, detect an incoming email that matches a rule the user set (“flag anything from the CFO”), and take the configured action — summarize, draft a reply, schedule a calendar block — without the user being present. Per Engadget's launch coverage, Google describes it as: “Gemini Spark runs on dedicated virtual machines on Google Cloud, so it can keep working in the background 24/7 even after you close your laptop or lock your phone.”

This is also the architectural reason Apple deliberately chose the opposite design path. Apple Intelligence runs on-device (with Private Cloud Compute for heavier tasks) — that approach minimizes the blast radius of a misfired action, since the agent only acts in the context of an open app. Spark's always-on model trades that privacy-by-design guarantee for genuine persistence. Google's bet is that knowledge workers will accept inbox surveillance in exchange for the time saved — but the privacy trade-off is real and worth naming.

The underlying model powering the Spark VM tasks is Gemini 3.5 Flash — Google's mid-size multimodal model with a 1.05M token context window. Flash's speed-to-cost profile makes it a sensible choice for ambient background tasks that may run many times per day across a large subscriber base.

Spark's first-party integration surface is Workspace-complete: Gmail, Google Docs, Sheets, Slides, Drive, and Calendar are all live at launch. Beyond the core Workspace suite, Spark extends into Chrome for web actions and Android Halo for mobile-native triggers. The result is that Spark can act across the full arc of a knowledge worker's day — reading emails, producing documents, scheduling meetings, organizing files, browsing the web — without leaving the Google ecosystem.

One integration detail that sets Spark apart from session-bound assistants is the dedicated email address. Google confirmed at I/O that each Spark instance has its own email address, meaning users can delegate tasks simply by forwarding a thread (“handle the follow-up on this”) or writing a task description as an email body. That reduces the activation cost of Spark to near zero for anyone who lives in Gmail — no new interface to learn, no app to open.

The vertical integration here is Spark's structural advantage over every competitor. Apple Intelligence can summarize a Mail thread; it cannot read a Sheets file and then draft a Doc citing that data. Microsoft 365 Copilot can do that within the Microsoft ecosystem — but Copilot is session-bound and enterprise-priced. Spark is the first product to combine always-on persistence with cross-Workspace synthesis at the consumer price tier. For teams that live in Google Workspace, this is the most meaningful personal productivity product Google has shipped in years.

Third-party app access in Spark uses the Model Context Protocol (MCP) — the same open standard that Anthropic introduced and Google has adopted as the interoperability layer for Antigravity 2.0. At I/O launch, three MCP partners are live: Canva for design tasks, OpenTable for restaurant reservations, and Instacart for grocery ordering. Google confirmed more partners would follow in the coming months but did not name them or provide a timeline.

The MCP architecture matters strategically. Spark's first-party integration surface is effectively unlimited (Google controls Gmail, Drive, Calendar, Chrome, Android) — but the third-party surface depends on partner adoption. Three partners at launch is a deliberately modest start. The question is whether the MCP ecosystem grows fast enough to make Spark meaningful outside the Google productivity stack. For comparison, Apple Intelligence supports third-party app intents via the App Intents framework across hundreds of apps — but none of those integrations are ambient.

For a deeper look at how the MCP tool surface and Antigravity SDK work for developers building their own Spark-class agents, see our Antigravity 2.0 desktop agent deep dive. The shared harness story is where the platform ambition becomes clearest.

Gemini Spark and Antigravity 2.0 shipped on the same morning, and they share the same agent harness infrastructure. Understanding the relationship reframes Google's agent strategy from “two products” to “a platform with two exposure surfaces.”

The Antigravity harness has three architectural layers. The first is goal persistence and task decomposition — Spark's ability to break a high-level request (“watch my inbox and summarize weekly”) into a set of repeating sub-tasks with trigger conditions, state recovery, and approval checkpoints. The second is VM execution — the dedicated Google Cloud VMs that give Spark its “24/7” property. The third is the MCP tool surface — the abstraction layer that lets Spark (and Antigravity 2.0 developer agents) invoke Canva, OpenTable, Instacart, and any future MCP partner through a standard protocol.

Spark is Google's consumer reference implementation — the product that demonstrates what the harness can do at the personal-assistant layer. Antigravity 2.0 is the developer surface: a standalone desktop app, CLI, Python / TypeScript / Go SDK, and Managed Agents API that lets external developers build their own ambient agents using the same primitives. The Antigravity 2.0 deep dive covers the developer harness in detail; this post focuses on the Spark consumer product.

The platform ambition is visible in the structure: Google is not just shipping a consumer agent — it is building the agent execution infrastructure and then showing developers what to build on top of it. That mirrors the pattern Google used with Firebase (consumer app → developer infra) and with Android (consumer OS → developer platform).

Google restructured its AI subscription tiers on the same day as the Spark announcement — a pairing that is not coincidental. The previous AI Ultra plan cost $250/mo. As of today, the top AI Ultra tier is $200/mo (a $50 cut) and a new $100/mo AI Ultra entry tier has been introduced. Per the official Google blog post from VP Shimrit Ben-Yair, the $100 tier provides 5× the Gemini-app usage limit of AI Pro; the $200 tier provides 20× Pro usage.

Both Ultra tiers include Spark access (when the beta opens week of May 25), 20 TB of Drive storage, YouTube Premium, and priority access to Antigravity 2.0 quotas. AI Pro ($19.99/mo) also gates Spark but at lower usage limits — the exact Spark usage cap at Pro tier has not been disclosed.

The repricing logic is worth reading carefully. Spark runs on Google Cloud VMs 24/7 — that is real compute cost. At $250/mo, the previous Ultra tier was the most expensive AI subscription from any major lab. Cutting to $200 (and adding a $100 entry) lowers the acquisition barrier for the Workspace-native audience Google needs to scale Spark. The $100 tier is a direct signal that Google wants to reach knowledge workers who view $250/mo as prohibitive — think SMB, freelancers, mid-market professionals.

For most Google Workspace users evaluating Spark, the $100 AI Ultra entry tier is the decision point. It unlocks full Spark functionality at a price that is lower than Microsoft 365 Copilot ($30/seat/mo enterprise, which typically requires a 365 Business Premium or E3 anchor license) while being meaningfully cheaper than the old $250 Ultra. The question is whether the Spark beta week of May 25 demonstrates enough ambient value to justify $100/mo compared to a $19.99 AI Pro subscription with lighter Spark access. That is a question for the complete Google I/O 2026 announcement guide follow-up to address once beta data is available.

The landscape comparison below is the most useful frame for understanding why Spark is a category move. Five vendors, five different product philosophies — the only persistent, ambient personal agent at launch is Spark. The Claude row is the sharpest editorial choice: Anthropic is not behind, it has made a deliberate product decision to focus on task-driven enterprise developer primitives rather than consumer ambient agents. Anthropic's May 19 announcements (Code with Claude London — self-hosted sandboxes + MCP tunnels) were enterprise-facing, not consumer-facing.

A note on the “blast radius” column: ambient agents that can send emails, schedule meetings, and create files on behalf of a user have a structurally higher blast radius than session-bound assistants. A misfired Spark task could send a draft to the wrong recipient, accept a meeting on behalf of the user, or delete files. Google's staged rollout (trusted testers before beta, US-only, Ultra-gated) reads as a deliberate blast-radius buffer. Apple Intelligence's on-device design keeps blast radius low by design — actions are scoped to open apps. Copilot adds confirmation steps for irreversible actions. Spark's approval-checkpoint model is still being documented.

Google's rollout sequence — I/O announcement → trusted testers same day → AI Ultra beta week of May 25 → AI Pro access TBD → international TBD — is unusually careful for a Google consumer product launch. The staging is intentional, and the reason is the blast-radius problem that is structurally unique to ambient agents.

A session-bound assistant like Apple Intelligence can generate a badly worded email draft, but it will not send that draft without the user explicitly pressing “Send.” An ambient agent with inbox access and email-send permissions can — depending on how the approval checkpoint is configured — send the wrong email to the wrong boss before the user is aware a draft was even created. The blast radius of a misfired ambient action is orders of magnitude larger than a misfired session-bound action. Google's Daily Brief and AI Inbox features took months to roll out carefully in Gmail precisely because of this risk.

The geographic constraint (US-only) is the second reliability signal. Google's international Workspace markets (EU, India, Japan) have different regulatory environments — GDPR in Europe creates specific consent and data-residency requirements that ambient inbox surveillance makes significantly more complex. Rolling out in the US first gives Google a contained blast radius while those frameworks are worked out.

The AI Ultra gate is the third reliability buffer. Ultra subscribers have self-selected as power users comfortable with Google's AI infrastructure and terms of service. They are more likely to configure Spark conservatively, read the approval settings, and report misfires in a way that helps Google tune the safety defaults before broader rollout.

For enterprise teams considering Spark, the immediate question is not “how powerful is Spark” but “what is the approval-checkpoint model and who controls it.” Google has not fully documented the default behavior of Spark's action-confirmation flow as of May 19 — that is the information to wait for before enabling Spark on team accounts. Our AI transformation engagements help teams evaluate ambient-agent capabilities and configure them safely, including Spark's action gates as documentation becomes available.

Three near-term developments will determine whether Spark is a category-defining product or a power-user niche. First, MCP partner expansion: the three launch partners (Canva, OpenTable, Instacart) cover design, dining, and grocery — three use-cases that are plausible Spark demos but not the primary workflows of a knowledge worker. The question is whether Google can sign travel, CRM, project management, and communication tools (Salesforce, Notion, Slack, Zoom) within the next quarter. Without those integrations, Spark's third-party surface is too narrow to displace session-based workflows for most professionals.

Second, the EU rollout timeline. GDPR Article 22 creates restrictions on fully automated individual decision-making — an always-on agent that sends emails or schedules meetings on behalf of users sits in a legal grey zone that will require explicit consent flows, data-residency controls, and possibly a DPA review. Google has not disclosed an EU timeline, and the US-first rollout suggests the EU framework is not yet resolved. Interpreting the timeline conservatively: EU availability in 2026 H2 if the consent framework moves quickly; 2027 H1 if not.

Third, the May 25 AI Ultra beta wave is the first real reliability signal. Trusted testers are by definition a biased sample — they will configure Spark carefully and report issues proactively. The week of May 25 is when the first paying subscribers who did not opt into a testing program start using Spark in production. Misfired actions at that scale will be the first genuine blast-radius data Google has to work with.

From a platform strategy lens, the most important signal is whether Antigravity 2.0 SDK adoption drives third-party agents that extend Spark's capability surface beyond what Google can build first-party. If the Antigravity harness becomes the go-to infrastructure for consumer ambient agents — not just enterprise automation — Google will have accomplished something more significant than shipping Spark: it will have defined the agent execution stack for the next generation of personal software. See Anthropic's MCP tunnel launch for the parallel enterprise developer story from the same day.