In the concrete experiment, a reporter ran Spark against his own Drive and Gmail to task the model with compiling

Independent hands‑on tests show Google’s agentic model Gemini Spark can autonomously find files in a user’s Google Drive, extract and compute figures from them, and assemble a personalized Gmail draft — demonstrating cross‑service automation that could materially change what assistants can do for account holders.

In the concrete experiment, a reporter ran Spark against his own Drive and Gmail to task the model with compiling an email to his wife summarizing their average monthly grocery spending in 2026. Spark identified the wife’s address despite the email lacking her first name, located the correct spreadsheet even though its filename did not include the word “budget,” extracted month‑by‑month totals (including partial May data), computed an average, and created a Gmail draft that included a personal sign‑off.

The tests mirror capabilities Google showcased in live demonstrations of Gemini Spark at last month’s I/O event. That contrasts with Apple’s approach, which so far has leaned on concept videos for its forthcoming Apple Intelligence and Siri enhancements rather than live stage features — a distinction critics such as John Gruber have repeatedly flagged when assessing Apple’s public demos.

For developers and integrators the practical takeaway is that agentic models can orchestrate multi‑step workflows across services (Drive data → computation → Gmail draft), disambiguate account identities, and act autonomously within a user’s account. If Apple’s planned Gemini‑backed Siri implements similar behaviors, users could see more robust on‑device or account‑level automations than current assistants provide.

Testers called the results impressive but cautioned they were not flawless. The reporter described being “floored” and called the output “scarily good,” yet also noted Spark did not fully deliver on every staged example. Those caveats suggest vendors will likely roll out agentic features in phases and that real‑world validation will be necessary before such behaviors are relied on in production applications or services.

The hands‑on experiments were summarized on Jun 2, 2026, in a piece by Ben Lovejoy that referenced original testing by Jay Peters of The Verge. Builders and observers should track the timing of broader availability, the scope of privacy controls, and any documented APIs as companies move from staged demos to public releases.