Gemini Spark: Google's 24/7 AI agent that wants to become your second brain

Gemini Spark: Google's 24/7 AI agent that wants to become your second brain

🔎 Google just launched the most ambitious agent of 2026

Google I/O 2026 was not just another event on the tech calendar. The presentation of Gemini Spark marked a turning point: for the first time, a tech giant is offering an AI agent designed to run continuously, 24 hours a day, 7 days a week, directly in the Google cloud. Not a simple extension of Gemini. A true autonomous system that lives in your ecosystem.

The announcement, picked up by TechCrunch and Mashable, immediately shook up the autonomous AI agent market. The reason is simple: no one had yet dared to connect a permanent agent to the entire Google suite with the ability to trigger financial actions.

The context is that of an intensifying agent war. OpenClaw has dominated the open-source segment since early 2025, Anthropic's Claude Code has taken control of developer workflows, and Google seemed to be lagging behind. Gemini Spark is the answer. Aggressive, integrated, and potentially disruptive.

This article breaks down the product, its real capabilities, its limitations, and what it concretely changes for a user or a business.

The essentials

• Gemini Spark is a cloud AI agent that runs 24/7, natively integrated with Gmail, Calendar, and Chrome.
• It can create specialized sub-agents to delegate complex tasks.
• Google is launching the "Agent Payments Protocol", a secure protocol allowing Spark to make payments on your behalf.
• The beta opens this week (May 2026) for AI Ultra subscribers in the United States.
• The underlying model is Gemini 3 Pro Deep Think, which achieves an agentic score of 95.4 on reference benchmarks.

Recommended tools

What Gemini Spark actually is — A cloud daemon, not a chatbot

Gemini Spark is not a new conversational interface. It's a continuous process running in the Google cloud that monitors, triggers, and executes actions without human intervention.

Concretely, you give it goals, not prompts. You say "Manage my travel bookings for the June conference and optimize prices," and Spark works in the background. It monitors prices, sends emails from your Gmail, creates events in Calendar, and only notifies you when a decision requires your approval.

The distinction is fundamental compared to Gemini 3 Pro in standard mode. Where classic Gemini answers a one-off question, Spark maintains a persistent state. It remembers your preferences, your budget constraints, your habits. It's exactly the type of capability found in AI memory architectures for agents, but here it is offered turnkey.

The engine is Google's Gemini 3 Pro Deep Think, which scores an agentic score of 95.4, behind only OpenAI's GPT-5.5 (98.2). A logical choice: this model excels in long reasoning and multi-step planning, two essential skills for an agent running continuously.

Gmail, Calendar, Chrome Integration — The Ecosystem Advantage Nobody Has

This is Spark's real trump card. Google has access to three of the world's most-used services: Gmail (1.8 billion users according to 2025 Google data), Google Calendar, and Chrome (65% desktop market share in 2026 according to StatCounter).

The integration goes far beyond simple reading. Spark can:

In Gmail: write and send emails on your behalf, sort your inbox based on complex rules, automatically reply to certain senders, extract information from attachments and cross-reference it with other data.

In Calendar: suggest time slots, manage conflicts, move meetings by negotiating with other people (via their email), anticipate travel times.

In Chrome: monitor web pages, fill out forms, extract data from online tables, and even interact with non-API web interfaces.

This depth of integration is impossible for a third-party agent like OpenClaw, which has to rely on browser extensions or limited APIs. This data access asymmetry constitutes Google's true strategic moat against the competition.

The parallel with Antigravity 2.0, Google's agent-first suite is obvious. Both products share the same philosophy: the agent is not a separate tool, it is the primary interaction layer with your data.

Sub-agents — Delegation and specialization

The most technically interesting feature of Spark is its ability to create sub-agents. You aren't handing over a monolith to manage. You are creating a hierarchy of specialized agents.

A concrete example: you are preparing a product launch. Spark creates a "competitive research" sub-agent that scrapes the web via Chrome, a "communications" sub-agent that manages email drafts in Gmail, and a "logistics" sub-agent that syncs Calendar with your team's availability. The main agent coordinates everything.

This architecture is strongly reminiscent of the plugin and extension system found in agents like Hermes Agent, where modularity allows for adding specific skills without overloading the main model.

The difference is that at Google, sub-agents automatically inherit the same access permissions to the ecosystem. No API configuration, no tokens to manage. The competitive research sub-agent immediately has access to Chrome, and the communications one to Gmail.

Google has not communicated a specific limit on the number of simultaneous sub-agents during the beta. But according to feedback from early testers cited by Mashable, cloud resource management is transparent — it is Google that absorbs the orchestration complexity.

Agent Payments Protocol — The agent that spends for you

This is the most discussed feature, and for good reason: Google allows Spark to make payments on your behalf via a new protocol called "Agent Payments Protocol" (APP).

The protocol works with three levels of security:

Level 1 — Automatic micro-payments: below a configurable threshold (default $10), Spark pays without confirmation. Ideal for recurring subscriptions, API data purchases, service fees.

Level 2 — Notification validation: between the low threshold and a medium cap (default $100), Spark sends a push notification with a transaction summary. You approve with a single tap.

Level 3 — Biometric validation: beyond the cap, fingerprint or facial recognition is required.

The protocol relies on Google Pay and is compatible with credit cards saved to your Google account. No crypto involved, contrary to the rumors circulating before the announcement.

What makes this system credible is its "opt-out" rather than "opt-in" design for micro-payments. The agent has permission by default to spend small amounts, which eliminates the friction point that kills the utility of autonomous agents. If an agent has to ask you for permission for every $3 action, it loses 90% of its value.

The comparison with the best autonomous AI agents is unequivocal: no competitor currently offers an integrated payment pipeline. OpenClaw can interact with payment APIs, but the configuration is entirely up to the user. At Google, it's native.

Gemini Spark vs OpenClaw — The duel that defines the market

The comparison is inevitable. On one side, Spark: proprietary, integrated, closed. On the other, OpenClaw: open source, self-hosted, infinitely customizable.

The choice is not binary. For a solo user or a small team living in the Google ecosystem, Spark offers a zero-friction experience that OpenClaw cannot match. You configure nothing, you delegate.

For a company handling sensitive data, with custom workflows, or that refuses to send its emails to Google servers for processing, OpenClaw remains the only viable option. The ability to run AI agents locally with Ollama or to choose your model among the best LLMs for agents is a structural advantage that Spark will never be able to offer.

In practice, the market will likely segment. Spark for the mass market, OpenClaw for power users and regulated companies.

Positioning relative to GPT-5.5 and Claude Opus 4.7

Spark is not a model, it's a product. But it is powered by a model, and this distinction deserves to be clarified.

Gemini 3 Pro Deep Think (95.4 agentic) is lower than GPT-5.5 (98.2) and slightly higher than Claude Opus 4.7 Adaptive (94.3) in terms of raw score. But the agentic score does not measure ecosystem integration. An agent with a score of 95 that has access to your Gmail, your Calendar, and your Chrome history will be more effective in 80% of daily tasks than an agent with a score of 98 that only has access to a chat window.

This is exactly the logic that led Google to bet on integration rather than pure benchmarking. Claude Opus 4.7 likely remains superior for pure mathematical reasoning or complex creative writing. GPT-5.5 dominates on coding tasks, which explains why developers continue to favor the best LLMs for coding like GPT-5.5 or Claude Opus 4.7 rather than Gemini for development work.

But for daily management — emails, scheduling, web searches, micro-decisions — Spark benefits from a huge structural advantage. Access to data beats pure reasoning in the majority of consumer use cases.

Availability and access — Who can use it right now?

The Gemini Spark beta opens this week (week of May 19, 2026) for AI Ultra subscribers in the United States. No public beta, no open waitlist. Access is strictly limited.

AI Ultra is the most expensive tier of the Google AI subscription, at $49.99 per month (May 2026, check on google.com). It includes access to the most powerful Gemini models, Gemini Advanced, and now Spark in beta.

Google has not announced an international rollout schedule. Historically, advanced Google features take 2 to 6 months to arrive in Europe, due to regulatory constraints (notably GDPR). The Agent Payments Protocol will likely add an additional layer of regulatory complexity, which could delay European availability.

For users outside the US who want to explore autonomous agents right now, the most realistic path remains to create an AI agent with open source tools or to use free AI APIs to build something similar, even if the ecosystem integration will inevitably be lower.

Implications for web hosting and creators

An agent that can browse the web, fill out forms, and make payments has direct consequences for the web hosting ecosystem and online content creation.

Spark can theoretically manage the entire process of putting a site online: purchasing the domain name, setting up the hosting, deploying the files. If you use a web host with a web interface (like Hostinger), Spark can interact with the control panel via Chrome.

This means that the time between "the idea for a web project" and "the project being live" could go from a few hours to a few minutes. The agent handles the logistics while you focus on the content.

For creators who manage sites or newsletters, Spark can also monitor performance, extract analytics from web dashboards, and alert you when a metric drops below a certain threshold. A level of automation that was previously reserved for those who knew how to configure scripts and webhooks.

Privacy Risks — The Faustian Bargain of the Google Ecosystem

The argument is well-known, but it bears repeating with precision. For Spark to work, it must read your emails, know your calendar, track your Chrome browsing, and have access to your payment methods. This is the highest level of permission any third-party software has ever requested from a mainstream user.

Google claims that the data processed by Spark is not used for model training. This is the same promise made for Google Workspace in general. But a promise is not a technical mechanism. There is no independent, verifiable proof that Spark's data remains isolated from the training pipeline.

The Agent Payments Protocol adds a financial dimension to the risk. A bug in level 1 (automatic micro-payments) could theoretically lead to unwanted expenses. Google has communicated about daily and per-transaction caps, but the technical details of the protocol have not been published as open source.

For users sensitive to these issues, two alternatives exist. The first is self-hosted OpenClaw, where your data never leaves your machine. The second is to run agents locally with Ollama, which eliminates any data transmission to a third-party cloud. The trade-off is the loss of Gmail/Calendar integration.

❌ Common mistakes

Mistake 1: Confusing Spark with Gemini Advanced

Gemini Advanced is an enhanced chatbot. Spark is an autonomous agent that runs continuously. These are not the same products, and they do not address the same needs. Advanced answers your questions. Spark acts without you asking any questions.

Mistake 2: Believing that Spark replaces code LLMs

Spark is not designed to replace Claude Opus 4.7 or GPT-5.5 in a development workflow. Its Chrome integration allows it to interact with web IDEs, but for serious code, the best LLMs for coding remain vastly superior. Spark is a general productivity agent, not a pair programmer.

Mistake 3: Underestimating the real cost

$49.99 per month for AI Ultra is the advertised price. But if you use Spark intensively with multiple sub-agents and continuous tasks, you consume Google cloud resources. The pay-as-you-go pricing model for background tasks has not been detailed, and it is likely that power users will exceed the initial plan.

Mistake 4: Enabling automatic payments without configuring thresholds

Level 1 of the Agent Payments Protocol is enabled by default. If you do not immediately configure the thresholds and caps, Spark can spend up to the default limit without consulting you. Take 5 minutes to adjust these settings before giving your first objective to the agent.

❓ Frequently Asked Questions

Is Gemini Spark available in France?

No. The beta is limited to AI Ultra subscribers in the United States (May 2026). No international rollout schedule has been announced. GDPR constraints could delay its arrival in Europe by several months.

Can Spark be used with a model other than Gemini 3 Pro Deep Think?

No. Unlike OpenClaw which allows you to choose your model, Spark is locked to Gemini 3 Pro Deep Think. You cannot switch to GPT-5.5 or Claude Opus 4.7, even via a third-party integration.

Is the Agent Payments Protocol secure?

The protocol includes three security levels and relies on the Google Pay infrastructure. However, the cryptographic details are not open source. For micropayments (<$10), no confirmation is required, which represents a risk of unwanted spending if the agent misinterprets your goals.

How many sub-agents can be created?

Google has not communicated a specific limit during the beta phase. Early feedback suggests that orchestration is seamlessly managed by Google Cloud, but a limit will likely be introduced at general launch for cost reasons.

Does Spark replace OpenClaw?

No. Spark and OpenClaw target different audiences. Spark is for users of the Google ecosystem who want zero configuration. OpenClaw is for developers and businesses who want total control, self-hosting, and the freedom to choose their model.

✅ Conclusion

Gemini Spark is Google's most ambitious product since Gmail in 2004. A 24/7 agent that lives in your email, your calendar, and your browser, capable of paying for you, had never been attempted at this scale. The bet is risky — both in terms of user trust and regulation — but the ecosystem advantage is real and probably insurmountable for proprietary competitors. For US users with an AI Ultra subscription, the beta is worth trying immediately. For everyone else, the time to seriously compare the meilleurs agents IA autonomes has arrived.