Google Gemini Spark Brings Always-On Agents To Workspace

Google used I/O 2026 to put a much clearer shape around its agent strategy. The Verge reported on 19 May that Google is launching Gemini Spark, an always-on AI agent that can work across Workspace apps, run in the background on Google Cloud and eventually connect to local files, Chrome, text and email. Google’s own I/O posts describe Spark as a 24/7 personal AI agent built on Gemini 3.5 and the Antigravity harness.

The product pitch is straightforward: Spark can write emails, prepare study guides, monitor card statements for hidden subscriptions, and keep working after a laptop closes or a phone locks. That last detail is the important one. Google is no longer talking about a chatbot waiting inside a tab. It is describing an agent runtime that lives close to a user’s documents, inbox, browser and app integrations.

The technical stack matters because it shows how quickly the agent market is becoming an infrastructure contest. Gemini Spark runs on Gemini 3.5, while Google says Gemini 3.5 Flash combines frontier intelligence with fast action. Sundar Pichai said Google is using 3.5 Flash internally with a reimagined Antigravity platform, processing more than three trillion tokens a day across AI developer tools after being at half a trillion in March.

That internal scale gives Google a feedback loop few companies can match. If Antigravity becomes the place Google tests agent orchestration at industrial volume, Spark becomes the consumer and knowledge-worker surface for the same underlying bet: agents need a harness, permissions, cloud execution, connectors and supervision, not only a clever model.

Google is also leaning into the Model Context Protocol. The Gemini app post says Spark is expanding connected apps with MCP links to Canva, OpenTable and Instacart, with more partners integrating. For builders, this is a strong sign that MCP is moving from developer-tool plumbing into mainstream product distribution. If Google normalises MCP as the way agents reach external systems, app vendors will feel pressure to expose useful, permissioned actions rather than another thin chatbot endpoint.

The release also raises the standard privacy and control questions. Google says Spark operates under the user’s direction and will ask permission before high-stakes actions such as payments or sending email. That is the right language, but the operational detail will decide whether the product feels useful or reckless. Persistent agents need transparent logs, revocable permissions, narrow scopes and obvious failure handling. A vague consent prompt is not enough when the agent can touch inboxes, documents, purchases and local files.

Antigravity 2.0 is the developer-facing half of the same announcement. Google says it is expanding the platform beyond coding into a standalone desktop application for managing cohorts of autonomous AI agents, alongside a command-line interface and an SDK for building AI tools. That puts Google closer to the full agent workbench category: plan, delegate, observe, intervene and ship.

The most interesting competitive point is that Google can bundle this through accounts people already use daily. OpenClaw and similar systems proved there is demand for persistent personal agents that can work through messaging apps, files and local context. Google’s advantage is distribution and integration. Its challenge is trust. The more Spark does in the background, the more users will need confidence that it is acting within clearly visible boundaries.

For teams building agent products, Gemini Spark is worth treating as a market marker. The baseline is shifting from conversational assistants to supervised background workers with real connectors. The next argument will not be whether agents can act. It will be which platform gives users enough control to let them.

---

Published: 2026-05-20 - Sources: The Verge article by Emma Roth published 19 May 2026; Google I/O 2026 keynote post; Google Gemini app announcement; Google AI subscription update.