Google DeepMind's Logan Kilpatrick: Why the Model Eats the Harness

Summary

Logan Kilpatrick, who leads Google AI Studio and the Gemini API, joins Sonya Huang to discuss Google's agentic AI strategy, the Antigravity agent harness, coding model competition, generative media, world models, Android app building, and DeepMind culture. He argues models will absorb external scaffolding over roughly 12 months, moving the edge away from standalone harness startups. He also says AI has been positive sum for Search, sees coding as narrow superintelligence, and expects vertical superintelligence in math, finance, and science before AGI. The conversation frames Google's integrated AI stack as a key competitive asset while highlighting opportunities for focused startups in vertical domains.

• Google is rebasing products around an agentic layer powered by Antigravity, intended as a common harness across Search, Gemini app, Cloud, and AI Studio.
• Logan argues AI has been positive sum for Search rather than cannibalistic, with people searching more even as agents do more.
• Google is pushing coding model competitiveness with Antigravity, pretraining resources, and Gemini 3.5 Flash despite current developer mindshare split with Claude and Codex.
• Omni is described as a true single model for text, audio, music, image, and video, with strong early video-editing use cases.
• AI Studio's Android app-building feature has produced 350,000 apps, many personal-use apps, positioning Android as a builder platform.
• Logan expects the model to absorb external agent harnesses over roughly 12 months, making standalone harness alpha less durable.
• He sees jagged vertical superintelligence arriving first in verifiable domains such as math, finance, and science, before AGI.
• DeepMind's culture is framed as scientific, fast-moving, and focused on deploying Gemini to billion-user products.