Gemini 3 combines one-shot UI generation with agentic tool calling for more complex software tasks.
2
Nano Banana Pro adds Google Search grounding, better text rendering, image consistency, localization, and creative controls.
3
Google AI Studio is moving from front-end prototypes toward full-stack applications whose databases, APIs, and other services can be configured through prompts.
Summary
Kat Kampf and Ammaar Reshi demonstrate how Gemini 3, Nano Banana Pro, and Google AI Studio change the way applications can be built. Gemini 3 can generate polished interfaces from a single prompt and handle more complex tasks through agentic tool use. Nano Banana Pro adds search-grounded image generation, improved text rendering, localization, consistency across people and images, and controls such as changing the focus of an existing picture. The demos include a branching comic book, personalized laptop stickers, a cloned AI Studio interface, and a 3D racing game. The presenters also show AI Studio's current limits. The multiplayer racing game supports more than 20 players in the lobby, but the race cannot start because everyone is blocked at the starting line. They describe upcoming backend support and a full-stack runtime that can set up services such as Express, storage, and payment integrations through prompts. Their goal is to make software creation accessible to people who do not begin with a traditional IDE.
Gemini 3 pairs strong interface generation with agentic tool use
Kat Kampf describes two capabilities behind Gemini 3. It can understand design and generate a good website or user interface in one shot. It can also call tools and work through complex tasks in large codebases. She places these uses on a spectrum: a developer may want a quick one-shot website, or may need an agent to operate across a more involved software project. She points to a software engineering experiment in which Gemini 3 performs above the other models shown in agentic scenarios and above Google's earlier models.
Nano Banana Pro uses search knowledge and renders image text more accurately
Ammaar Reshi presents Nano Banana Pro as an image model powered by Google Search. A user can ask how to make a tea, for example, and the model can search for information before creating an infographic or diagram. Reshi also focuses on text rendering, showing text wrapped around a can and translated into another language on the same reference image. He says the model supports many languages and gives more consistent results, including group images with up to 14 people. A simple instruction can also change the focus from a person to flowers while preserving the rest of the image.
AI Studio lets users build and share AI applications without managing every API bill
Google AI Studio provides access to Gemini models, API keys, a chat interface, and a prompt-to-application build experience. Its AI chips expose tools such as Google Search grounding, Google Maps grounding, and the Live API. Kampf gives examples of a tennis-swing app that provides live corrections and a posture app that calls out when someone leans forward. For most models, users can build and share an application without adding their own API key. Visitors to a shared application use their own AI Studio free quota, so the creator does not have to absorb every use as an unexpected API charge.
Nano Banana Pro makes generated stories more useful through readable text and consistent characters
Reshi uses Nano Banana Pro to make a personalized comic book. Users upload faces, choose a story genre and language, and describe an adventure. The demo places Reshi and Kampf in a story about presenting AI Studio at AI Engineer in New York. The generated comic includes readable story text, recurring characters, background details such as an AI Engineer banner, and a branching choice partway through the narrative. Reshi says the model's consistency makes this kind of use case possible, while Gemini 3 supplies the story generation and the application lets users change the direction of the story as it is being created.
Gemini 3 can produce more considered visual designs from simple prompts
The presenters show a website generated from the instruction to create a slick animation website. The result uses shader animations, page transitions, effects, and typography selected by the model. Reshi contrasts this with the purple gradients he associates with earlier generated websites. He says people who struggled with design or tried to work around Figma can prompt their way toward something visually polished. The point of the demo is that the design result comes from a short prompt rather than a detailed manual design process.
Search grounding turns personal sticker generation into a context-building task
Kampf's laptop-sticker app can use words supplied by the user or search Google for information about a person. She enters Ammaar Reshi's name and selects a 1K resolution. The application retrieves current sources, builds context about his interests, and generates stickers based on that information. The output includes his children's book, Alice and Sparkle, which Kampf says is represented because he has discussed it frequently. She also says the same approach can use information about current news events rather than relying only on a model knowledge cutoff.
AI Studio can help its own team explore product ideas before implementation
Kampf shows an application made by prompting AI Studio with a screenshot of AI Studio itself. The instruction asks it to clone the interface closely and add an export flow to Google's Antigravity IDE. The generated app recreates the light-mode interface and adds an Antigravity button. Kampf says the model sometimes invents different export patterns, such as a command-line interface or an export status view. She sees this as a way to explore product interactions without being limited to existing patterns in the product being designed.
The planned full-stack runtime will hide infrastructure choices behind application intent
Reshi shows a 3D racing game generated as a front-end React application using 3JS. He says the next step for AI Studio is backend support and a full-stack runtime. A user should be able to ask for a multiplayer application without specifying that it needs Express or wiring the pieces together manually. In the live test, more than 20 players join the lobby, but collisions leave cars stuck near the starting line and the race cannot begin. Reshi presents this as a working demonstration with visible limits, while describing future support for first-party and third-party APIs, storage, and payment services.
