Gemini Spark now runs recurring tasks on your photos

Google's personal agent got write access to a data store this week, and it picked the one with the highest emotional value and the lowest reversibility: your photo library. Gemini Spark can now edit images, build albums, share them out through Gmail, and — the part that matters — run recurring tasks in the background without you in the room.

What actually happened

TechCrunch reported the rollout on September 4, after Google Photos lead Shimrit Ben-Yair announced it. Per Google's own support documentation, Gemini Spark can search and organize photos by subject, location, date, or event; enhance images and generate collages; create and organize shared or private albums; extract text from images; distribute album links through Connected Apps like Gmail and Messages; and set up recurring tasks — Google's example is automatically creating a monthly highlight album.

It is limited: Gemini AI Pro and Ultra subscribers, United States, English, 18 or older, mobile and web only, and gradual rollout — so it may not appear for you yet. TechCrunch notes Google did not say whether or when it goes international.

One design detail is worth flagging: per Google's docs, edits create new copies and the originals stay unmodified.

Why agent permission scoping matters for your business

Ignore the photos. Look at the shape.

Google just shipped a consumer product where an agent holds standing write access to a personal data store, runs on a schedule, and can push content outward into email and messaging. That is precisely the architecture you will be asked to build over your CRM, your inventory table, or your invoicing system within the year — and the questions Google had to answer are the ones you will have to answer.

Google answered the hardest one with a copy-on-write rule. The agent can edit anything; it cannot destroy anything, because an edit produces a new object and the original stays put. That single constraint turns "the agent did something weird to 400 photos" from an incident into an annoyance. Build the same guarantee into your own agents: append, never overwrite. New row, new version, new draft — with the prior state still queryable.

Then handle the other two. Scheduled is different from prompted: an agent that fires at 3am with nobody watching needs a tighter permission set and a delivery report, not the same token you use interactively. And outbound is its own category — the moment an agent can send email on your behalf, its blast radius stops being internal. Gate that path separately, and keep a log of every message it sent, addressed to a human who reads it.

Standing access, on a schedule, with an outbound channel. Get those three scoped before you get the demo working, not after.

Sources: TechCrunch, Google Gemini Apps Help.