Lilith.
⌕
Editorial illustration: Google makes Tasks free, while the agent reaches into calendars, passwords and payments
Lilith illustration · editorial remix

Google has renamed its Spark agent to Tasks and started a free rollout in the US, Japan, Korea, Canada, Australia and India. The feature runs on Gemini 4 Hydrogen, with a home screen switch in the Gemini app between chat and task modes.

Paid Spark becomes a free layer inside Gemini

Spark had been available to Google AI Pro and Ultra subscribers since May. Tasks now brings its multistep work to hundreds of millions of people at no extra charge. The agent can build meal plans, manage calendars, find recurring expenses in email and work through connected web services.

The scope differs by country. Advanced features begin in the US, where Gemini can call businesses about prices and stock, sign into supported accounts through Google Password Manager and help complete purchases with Google Pay. US users can also assign work by email or text message.

A free agent is a distribution bet on the entire Google account

The main advantage of Tasks is not the model alone. It is access to email, calendars, files, photos, passwords and payment cards that users already keep with Google. The company can turn chat into action without asking people to rebuild context across several applications.

That creates a high barrier for competitors. An agent with a better benchmark still needs the permissions and integrations Google can offer inside its own account. For users, the cost of error rises too. A poor summary is annoying, while a wrong booking or purchase reaches into the real day and wallet.

Sandboxing limits damage, but mistakes remain

Google says each task runs in a hardware isolated virtual machine dedicated to one account. It adds prompt injection filters, Google Safe Browsing and confirmation before sensitive actions such as clicking Buy. Users can set routine actions to Always allow.

The company also acknowledges that Tasks can fail to finish or miss the instruction. Websites often block agents, so Gemini may hand the screen back before the final step. The security architecture reduces some risks, but it cannot fix a misunderstood intention or an error in source data.

Human intervention rates will show whether Tasks saves time

Useful evidence will include completion rates, corrections, cancelled purchases and the number of times the agent hands work back to a person. Differences among countries matter because the most attractive channels and payment features are initially limited to the US.

Google also needs to show that people can understand permissions across more than 60 connected apps. The agent becomes a useful assistant only when users can see both the result and the route it took to get there.

Lilith's verdict

Google is not merely giving people a free agent. It is handing that agent keys to the inbox, calendar and wallet, so the costliest mistake will be a wrongly executed task rather than a bad sentence.