Google introduces Gemini Spark, a 24/7 personal agent

Background

A chatbot waits to be asked. An agent is software that is given a goal and then works through the steps itself, with the person supervising rather than driving every move. Google used its I/O developer conference to put that idea at the centre of its announcements, in the form of a personal agent meant to keep working between conversations.

What Google announced

At Google I/O on 19 May 2026, Google introduced Gemini Spark, which it described as a personal AI agent that runs 24/7 and acts on the user's behalf under their direction^[1]. The agent was built on Gemini 3.5 Flash, a model Google announced the same day^[1].

Running 24/7 is the distinguishing claim: the agent is not tied to a single session or a single question, but is meant to be available continuously and to take actions for the user rather than only produce text for them to read.

How it rolled out

Spark was not available to everyone at once. It went first to trusted testers, with a beta for US subscribers to Google AI Ultra planned for the following week^[1]. Google's release notes described Spark as being in Beta for Google AI Ultra subscribers aged 18 and over^[2].

On 30 June 2026, Spark arrived in the Gemini macOS app^[2].

The rest of the same day

The launch came as part of a wider set of changes. Google merged Gemini Live into conversations, launched Gemini Omni for video creation, and introduced a Google AI Ultra tier starting at $100 a month^[2].

Google also said Gemini 3.5 Flash had become the default model in the Gemini app and in AI Mode in Search globally^[1]. It said the model generates output tokens 4 times faster than other frontier models^[1]. And it said Gemini 3.5 Pro was already in internal use, with rollout planned for the following month^[1].

Those details matter for reading the Spark announcement. An always-on agent spends much of its time on ordinary, high-volume work: reading pages, checking state, drafting updates, repeating small steps. Speed per token and cost per call decide whether such an agent can run continuously for a large number of people. Making the faster model the default in the main Gemini app and in Search put it in front of the general audience as well as the agent's early users.

Why it matters

Three things about the release are worth noting.

First, the shift in what the product is. A personal agent that acts on the user's behalf changes the interface from question-and-answer to delegation: the user states a direction, and the system carries out steps. That raises questions the chatbot frame did not — what the agent is allowed to do without asking, how the user reviews what it did, and how a process that runs continuously is paused or stopped.

Second, the way it was gated. Spark was in beta, limited to Google AI Ultra subscribers aged 18 and over^[2], and that tier started at $100 a month^[2]. Continuous, action-taking agents were thus offered first as a premium feature to US subscribers rather than as a default part of the free product.

Third, the timing. The agent, the model underneath it, the merging of Gemini Live into conversations and the video-creation model Gemini Omni were all announced or launched on one day^[1][2], and the macOS app followed about six weeks later^[2]. For readers looking back, the useful markers are those dates: 19 May 2026 for the announcement and the model, 30 June 2026 for the desktop arrival.