Vibe coding to production: Logan Kilpatrick on the evolution of AI Studio
Summary
In this live floor interview at Google Cloud Next '26, Sam Witteveen (Google Developer Expert) speaks with Logan Kilpatrick, product lead for Google AI Studio at Google DeepMind, about the rapid evolution of vibe coding, AI agents, and the Gemini platform ecosystem.
Kilpatrick opens by characterizing the conference's theme as the arrival of the "era of agents," noting that previous Google Cloud Next events felt speculative whereas this one demonstrates agents actually delivering on their promise. He describes AI Studio's journey through several distinct eras: starting as a prompt-to-prototype tool for getting API keys, then crossing a chasm roughly 18 months prior to actively helping users ship production applications, and now becoming a full vibe coding platform.
The conversation covers several new AI Studio features in depth. The "Build Tab" (launched at Google I/O the previous year) lets users go from a text prompt to a deployed application on Cloud Run in minutes, often at little to no cost. New additions include "Design Previews," which shows multiple UI theme options during generation; an "I'm Feeling Lucky" button that generates app ideas for users who don't know where to start; "Tap Tap Tab," a generative autocomplete feature powered by Gemini Flash that helps users articulate complex app ideas as they type; and "Yap to App," a voice-to-application feature that Kilpatrick describes as the second most popular build method, where Gemini actively reformulates loosely spoken ideas into coherent prompts before building. An "Edit Mode" enabling users to draw on app previews or select UI elements for real-time AI regeneration was described as imminent.
Kilpatrick addresses the shift from "vibe coding" to "agentic engineering," sharing that Google's own internal teams now use this workflow to make changes to AI Studio itself — with a structured handoff model where agentically engineered code gets reviewed, tested, and merged by senior staff engineers. He frames this as a productive partnership model that enables new contributors while maintaining production quality standards.
On the topic of agent capabilities, Kilpatrick and Jeff Dean (mentioned in passing) discussed how long-running agents are evolving from hours-long to potentially days- or weeks-long autonomous operation. He highlights the Gemini Live screen-sharing feature as enabling "omnipresent" tutoring and support use cases, citing a customer example where the live model walks non-technical users through complex product UIs in real time.
Kilpatrick draws a direct analogy between the current moment and the YouTube era: just as YouTube democratized content creation, AI is democratizing software creation. He argues this is net positive for traditional developers too, as the total addressable market for software expands and more builders will need technical partners. He cites a non-coder on his own team who became AI Studio's second-highest internal token consumer by building useful tools without writing a single line of code.
On the model side, he notes consolidation of specialized generative media models (Nano Banana for images, VEO for video, Lyria for music, Live for streaming audio/video) into the mainline Gemini architecture is the direction of travel. He previews significant investment in coding models, mentions Gemma 4 for on-device inference as exciting, and flags robotics as roughly 12–18 months from major breakthroughs. He closes with a call for user feedback and frames AI Studio's mission as urgently deploying technology so society can engage with it rather than theorize about it.
Key Takeaways
- •AI Studio's Build Tab enables prompt-to-deployed-app in minutes on Cloud Run, often free — millions of users are already building this way.
- •The 'Yap to App' voice feature is the second most popular way to build in AI Studio; Gemini actively reformulates garbled speech into a coherent app spec before building.
- •Adopt an 'agentic engineering' partnership model: vibe coders generate code agentically, senior engineers review and own it — this lets non-coders contribute to production systems while maintaining quality.
- •Raise your ambition ceiling: the models have crossed a threshold where you can ask for 30 things at once instead of one — the bottleneck is now the builder's imagination, not the model's capability.
- •Long-running agents will shift from hours to days or weeks of autonomous operation in the next 12 months — start designing workflows that assume minimal human-in-the-loop.
- •Gemini Live screen-sharing unlocks omnipresent help use cases (real-time UI guidance, debugging, tutoring) that represent significant untapped product alpha.
- •If you haven't tried a capability in the last two weeks, try it again — model improvements compound rapidly and prior failures may now work perfectly.
- •Google is consolidating its generative media models (images, video, music, live audio) into the mainline Gemini architecture — expect a simpler, unified model surface.
- •Democratizing software creation is net positive for professional developers: more builders means more demand for technical co-founders and senior engineering partners.
- •Token supply constraints are real even at Google's scale — identify your highest-leverage AI use cases now rather than applying AI indiscriminately.
Transcript
Hello and welcome to Google Cloud Next and I'm joined by Logan Kilpatrick. Logan, great to have you here. Thank you. I'm excited. This is This is a super cool setup. It's awesome to be like literally live in the venue right now. >> Exactly. We're sitting right with hundreds of people around us and stuff like that. Um what uh walking around, what have you seen today that that's been interesting? Yeah, it feels like the the era of agents is upon us. I think this morning's keynote you heard from fr...
