Gemini Spark is a personal AI agent announced at Google I/O 2026 that runs on Gemini 3.5 and is supported by Google’s Antigravity harness. The agent operates continuously in the cloud on dedicated Google virtual machines and does not require a phone to be unlocked or a laptop to be open, allowing it to run persistently without direct device interaction.
At launch, Spark works natively with Gmail, Google Docs, Google Slides and Google Workspace. Starting at launch, Spark gains MCP connections to Canva integration, OpenTable connection, and Instacart. Those third-party connections are available immediately at launch.
The Google I/O announcement presented Spark as a continuously running cloud-based personal AI agent. Launch materials emphasized native Workspace ties and the inclusion of third-party MCP connections.
Gemini Spark can automatically scan monthly credit card statements to identify new or previously hidden subscriptions, highlighting recurring charges and new services for user review. The agent can monitor a child’s school emails, extract deadlines and other schedule information, and compile that information into a daily digest that can be sent to a caregiver or guardian. Spark can extract raw notes from meetings across Gmail and Docs, convert those notes into a clean, organized document, and send a follow-up email to kick off the next phase of work. Users can teach Spark custom skills by programming their preferences, allowing tailored automations based on explicitly defined rules.
Google states Spark may do things like share your info or make purchases without asking. Users are advised to supervise Spark and grant permission before high-stakes actions.
Users can teach Gemini Spark custom skills by programming their preferences, enabling the agent to act according to explicitly defined rules. Over the summer, Spark will be able to receive texts and emails sent to it directly, and users will be able to create custom sub-agents to handle specialized tasks. Google plans to allow Spark to operate the local browser as part of its upcoming capabilities, expanding the agent’s ability to interact with web-based workflows on behalf of users. These customization paths are described in launch materials as user-programmable mechanisms that adapt Spark’s behavior to individual preferences.
A macOS desktop version is planned to extend Spark to local files on users’ machines. Internal references in beta or leaked materials included the codenames “Gemini Agent” and “Remy.” Google warned that Spark may share user information or make purchases without asking, and it advised users to supervise the agent’s activity. The company recommended that users grant permission before permitting high-stakes actions.
The announced roadmap links programmable customization with expanding communication and local-file access while pairing those capabilities with explicit privacy and permission cautions. Deployers and users are advised to consider supervision and permission settings when enabling Spark.
Gemini Spark is presented as a cloud-based personal AI agent with extensive integrations and automated features that run continuously in the background to assist users across email, documents and other workflows.
Decrypt currently has limited access and could not test the agent, so firsthand evaluation of its behavior and controls was not available at the time of reporting.
