Google’s Gemini Spark Tests the Limits of Always-On AI Agents

Google has placed a sizable bet on AI that doesn’t wait for instructions. At its I/O 2026 conference, the company unveiled Gemini Spark, a persistent agent built to handle multi-step work across email, documents, calendars and the wider web. Unlike chatbots that respond once and stop, Spark runs in the background on Google Cloud virtual machines. Close your laptop. Lock your phone. The agent keeps going.

TechCrunch reported the announcement in detail. Alphabet CEO Sundar Pichai told reporters it serves as “your personal AI agent that helps you navigate your digital life, taking action on your behalf and under your direction.” The system powers itself with Gemini 3.5 Flash and a framework from Google’s Antigravity project. Josh Woodward, vice president of the Gemini app and AI Studio, offered a vivid picture. “Even when you close your laptop or turn off your phone, Spark can keep working in the background as you go through your day,” he said. “When you use it, it almost feels like you’re tossing things over your shoulder, Spark’s catching them, and gets the job done.”

Early access began with trusted testers this month. A broader beta opened to Google AI Ultra subscribers in the United States, those paying roughly $100 monthly, and select business users. Official product pages confirm availability will expand quickly. Connections to Gmail, Calendar, Drive, Docs, Sheets and Slides sit turned off by default. Users must activate them. The agent does not scan inboxes indiscriminately. It acts only on assigned tasks, schedules or skills.

Practical examples abound. Spark can scan email threads for action items and build a prioritized to-do list. It logs receipts from group-trip planning into a spreadsheet then emails the updated plan. For freelancers it extracts client details from inquiries, creates a tracker row and sets up a dedicated Drive folder. Homeowners receive reminders about upcoming maintenance drawn from past invoices. The agent even browses multiple sites, compares options and assists with bookings. These demonstrations come straight from Google’s own overview page.

Third-party reach grows through the Model Context Protocol, an open standard. Planned connections include Canva, OpenTable, Instacart, Spotify, Expedia and Adobe. Later this year the agent will touch local files on Mac computers. The Verge noted the parallel to open-source efforts such as OpenClaw. Google positions Spark as its polished, cloud-native counterpart. Users will eventually text or email the agent directly, much like those experimental systems.

Real-world tests reveal both power and persistent gaps.

WIRED gave the beta a thorough workout. Writer Reece Rogers granted Spark access to his Gmail, Docs and Calendar then asked it, in one vague sentence, to plan his birthday party. The agent delivered a five-page itinerary within minutes. It located an actual karaoke reservation, compiled a guest list capped at the room’s capacity, outlined venue rules, suggested dinner options and drafted follow-up emails. The list ranked Rogers’ live-in boyfriend as a “close friend and frequent companion” yet omitted Rogers himself from his own celebration. “Your travel history and emails identify [partner’s name] as a close friend and frequent companion, making him a natural first addition,” the agent explained, according to the WIRED article published today.

The episode produced moments of sharp irony. Spark chose specific queer bars based on exact keyword matches in past itineraries and receipts rather than any inferred understanding of identity. When Rogers tested email drafting, the agent mimicked his casual tone but still required manual fixes for reservations that glitched during remote browser sessions. It triggered verification codes yet failed to complete bookings. Rogers ended up calling the sushi restaurant himself. The review captures a core tension. The output looked genuinely impressive. The relational blind spots and occasional hiccups left the author “wheezing with laughter” and uneasy.

Google stresses user oversight. The agent checks before major moves such as sending emails or initiating payments. Users set repeating schedules, upload custom skills that learn email style from past messages, or interrupt at any time. Even so, security warnings appear prominently. A malicious prompt could instruct the agent to extract private data and post it publicly. The company advises caution with full inbox access. Prompt injection remains a known risk across agent platforms. Rogers echoed that concern. Giving an AI unfettered context about one’s life carries consequences.

Recent coverage reinforces the stakes. DataCamp’s analysis from May 20, 2026 highlights how Spark chains tasks across Workspace while running on dedicated cloud resources. It contrasts the agent with reactive chatbots from competitors and flags the need for strong governance as capabilities expand. Enterprise versions tie into broader Google Cloud agent platforms, offering pre-built solutions and custom model training. Business users gain extra controls around data residency and audit logs that consumer Ultra subscribers lack.

Analysts watch closely. If Spark delivers consistent value on routine work such as inbox triage or trip coordination, adoption could accelerate. Persistent glitches or privacy incidents would slow it. The architecture marks a shift from on-device or session-bound assistants to cloud-resident workers that operate independently yet remain accountable. Google claims the agent learns from behavior over time through skills and schedules without constant retraining.

Executives avoid overpromising. Pichai and Woodward describe an active partner rather than an autonomous replacement. The system still needs clear direction. It excels at synthesis and execution but struggles with implicit social cues or common-sense adjustments. That gap showed plainly in Rogers’ party plan. The agent knew the maximum headcount for the karaoke room. It missed that the host belonged on the list.

Developers receive new tools. Antigravity now includes a desktop app, command-line interface and SDK for building, testing and deploying agents. These additions aim to let companies create domain-specific versions that sit alongside the consumer offering. Integration with the wider Google Cloud portfolio suggests Spark could evolve into a platform rather than a single product.

Privacy remains the largest open question. Deep access to email, calendars, documents and soon local files hands Google an unprecedented view of daily routines. The company says connections require explicit opt-in and that the agent does not read mail 24/7. Yet the always-on design invites scrutiny. Regulators and users alike will test whether controls match the ambition. Early feedback from testers, including the WIRED hands-on, suggests the experience feels magical when it works and jarring when it doesn’t.

Google plans rapid iteration. More MCP partners arrive over summer. Local file support on Mac lands soon. Enterprise preview expands. Consumer rollout widens beyond the Ultra tier. Each step brings the agent closer to the vision Pichai outlined: an AI that handles the mundane so people focus on what matters. Whether that future arrives smoothly or through a series of instructive stumbles will define the next chapter for personal computing.

One thing looks clear. The era of agents that simply answer questions has ended. The agents that act, persist and occasionally surprise their owners have begun.