Google's Gemini 3 learns to show its work, not just tell it

Google just released Gemini 3, and it's doing something its predecessors didn't: deciding how to answer you before it answers.

Instead of defaulting to a wall of text, Gemini 3 now looks at what you're asking and chooses the best way to respond. Ask for travel recommendations, and it might build an interactive webpage with images and follow-up questions. Ask for a diagram, and it sketches one. Ask for an animation, and it generates that instead. The model essentially picks its own format based on what would actually help you understand the answer.

Google calls this "generative interfaces," and it's a small shift that matters. Most AI tools force you to work around their constraints — you ask a question, you get text back, and you translate it yourself into something useful. Gemini 3 flips that. It does the translation for you.

The agent that actually gets things done

The upgrade also brings an experimental "Gemini Agent" feature that can connect to your Google Calendar, Gmail, and Reminders to handle multi-step tasks. You could ask it to "find a time next week when I'm free and book a dentist appointment," and it would break that into smaller steps, show you its progress, and pause for your approval before actually sending anything.

This is different from AI that just answers questions. This is AI that can coordinate across your actual life — at least the parts of your life that live in Google's ecosystem.

Where it shows up

Google is rolling these capabilities into its existing products. In Google Search, early users can now ask Gemini 3 Pro to generate deeper, more thorough summaries of topics instead of just listing links. For shopping, the model can now pull from Google's 50+ billion product listings to create interactive recommendation guides — think Wirecutter, but generated on the fly.

There's also "Antigravity," a new development platform that lets you describe what you want in plain language and have it generate code, tools, and workflows automatically.

The practical question now is whether this actually works at scale. Derek Nee, CEO of Flowith (an AI application company), noted that the speed and cost advantages are real, but "we need deeper testing to understand how far it can go." That's the honest take — the capabilities are there, but real-world performance across millions of users is still being written.

What's worth watching is the shift in philosophy. Instead of building AI that answers questions, Google is building AI that understands context well enough to choose how to communicate. That's less flashy than a bigger model, but it might be more useful.