Google is expanding Gemini Tasks to general availability, and the product has now been renamed three times: it started as Spark, became Task Mode, and ships as Tasks. The renaming is the least interesting part. The interesting part is that the assistant will now place phone calls and complete purchases.

What it actually does

Users can switch between Chat and Tasks inside Gemini and submit a request by text, SMS or voice. From there the agent can call a business on your behalf, book a service, pay through Google Pay, or log into a website using credentials you have already authorised.

Gemini's head, Josh Woodward, said publicly that he had used it to book vehicle maintenance and to analyse an electricity bill. Both are mundane tasks that are surprisingly hard to automate, which is exactly why they are good demos: a phone tree is a worse interface than a chat box, and a utility bill is a PDF with no schema.

The ecosystem is the moat

Google's advantage here is not the model. It is the plumbing: Search, Chrome, Google Pay, the account system, and integrations with Keep, OpenTable and Instacart. An agent that can read your calendar, find a restaurant, and pay for a booking without leaving the account is a different product from a chatbot that tells you to go do it.

That is also why this is harder for competitors than it looks. OpenAI can build a better agent, but it cannot inherit a decade of logged-in sessions across services it does not own.

Where the human stays in the loop

Google's design keeps payment behind an explicit confirmation, and the agent does not get your card number — it gets a token. Sensitive actions are handed back to the user rather than completed silently.

There is also prompt-injection defence, which is now the standard worry for any agent that reads web pages. A page can contain instructions, and an agent that follows instructions found in a restaurant menu is a security incident waiting to be written up.

The trust problem nobody has solved

An agent that books on your behalf will eventually book wrong: the wrong date, the wrong location, the wrong restaurant. The question is what happens next — whether there is a confirmation step, a cancellation path, and a record you can point at.

Google's answer is confirmation at payment and, implicitly, that the stakes of a restaurant booking are low. That works for consumer tasks. It does not generalise to anything with real money or real deadlines, which is why the enterprise version of this will look different.

What to watch

Whether Tasks becomes the default entry point rather than a mode. If it does, Google has effectively replaced the search box with a task box, and the advertising business has to be re-thought around that.

Whether the third-party integrations deepen. Booking a table is a partnership; rescheduling a flight is a negotiation with an airline's systems that nobody has automated well.

And whether Google publishes reliability numbers. An agent that completes 70% of tasks and silently fails the rest is worse than no agent, and no vendor has been willing to put a number on this yet.

The practical read

For consumers, this is a genuinely useful feature in a narrow band: bookings, appointments, bills. For anyone building products, the lesson is about the interface — the agent's value comes from owning the context, not from being smart.

If your product's workflow depends on a user manually re-entering information into someone else's system, that workflow is now a target. Not because Google will do it better, but because the expectation of not doing it has changed.