Jun 11, 2026, 8:00 AM ETInterview, Fireside Chat, Podcast

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

Google has officially entered an "agentic era" driven by the Gemini 3.5 model, moving beyond simple API interactions to products that take autonomous actions on behalf of users.
The "anti-gravity" agent harness serves as the new unifying through-line for Google's product ecosystem, replacing the Gemini API as the primary integration layer across hundreds of products.
Anti-gravity is designed as an ecosystem of tools including a core IDE, CLI, SDK, and web-based agent-first experiences to meet developers at various proficiency levels.
The same anti-gravity harness powers agentic capabilities in Search, the Gemini app, Cloud, and AI Studio, ensuring consistent agent behavior across the platform.
Google distinguishes between the base agent harness (80% shared functionality) and specialized versions optimized for specific use cases like "vibe coding" or 24/7 consumer assistance.
Sundar Pichai's characterization of the current period as the "agentic Gemini era" marks a shift from earlier Gemini 2.0 iterations which were considered premature for full agentic deployment.

Contrary to fears that agentic AI would reduce search volume, early data indicates a "positive sum" outcome where both human and agent-driven search activity have increased.
Google's strategic goal has shifted from maximizing "eyeball time" to maximizing customer outcomes, prioritizing task completion over user retention duration.
While some products like the Gemini app (24/7 agent) and Anti-Gravity (autonomous coding) are in "walk/run" phases, the majority of Google's 13 billion-user base remains in a "crawl" phase requiring human oversight.
The company anticipates that while enterprise adoption outside of coding is lagging, long-running agent tasks will see rapid improvement and adoption within the current year.
Google aims to avoid "cannibalization" by allowing agents to handle complex tasks, freeing users to pursue more ambitious problems rather than reducing total time spent on digital platforms.

The "vibe coding" phenomenon is accelerating, with teams launching mobile apps (e.g., Gemini macOS app) significantly faster than previous benchmarks due to agentic coding capabilities.
Google introduced "3.5 Flash," a post-training model that reportedly outperforms all previous Pro models in coding capabilities despite not involving new pre-training runs.
Internal metrics show a massive surge in token consumption within Google's own engineering teams, validating the "model eats the harness" theory where scaffolding becomes native to the model over time.
The "Avengers of AI" team at Google is currently focused on pushing coding frontiers, aiming to enable developers to build full-stack applications with minimal manual intervention.
Despite the "Code Red" narrative around competitors like Claude, Google's internal adoption and user feedback suggest a competitive landscape where 50/50 splits in developer tool usage are becoming common.

Google launched "Omni," a single unified model capable of generating and editing text, audio, image, and video, replacing the historical reliance on a suite of separate specialized models.
Omni's capabilities were demonstrated live when a crowd member's photo was edited in real-time to include a dog on stage, with the AI correctly synthesizing subtle interactions and reactions from other guests.
The definition of "world models" is evolving; they are no longer strictly action-conditioned video models but are now understood as systems with deep world understanding capable of diverse generative outputs.
Current limitations include the model not yet being state-of-the-art for all media types (e.g., audio/music) compared to specialized legacy models like Lyria or MusicFX, but these gaps are being addressed in future iterations.
Generative media is being framed not as replacing human identity (avatars) but as amplifying human content by modifying non-personal elements (backgrounds, props) while preserving the creator's voice and presence.

Google is observing a trend where "world models" will blur into "coding agents," requiring a mix of generative capabilities and traditional engine scaffolding (like game engines) for complex tasks such as video game creation.
The most popular categories of apps built in AI Studio have shifted from games (20% initially) to finance/crypto tools, personal productivity, and generative media.
Android is emerging as a primary platform for AI builders, with over 350,000 apps built in AI Studio in the last week, many of which leverage native OS capabilities unavailable to web-based agents.
The "Model Eats the Scaffolding" hypothesis suggests that external tools like search and code execution will eventually become native to the model, reducing the need for third-party harnesses over the next 12 months.
Future "superintelligence" is expected to first appear in verifiable domains like math, finance, and science, rather than open-ended general tasks, creating "jagged" rather than uniform intelligence.

DeepMind's culture is defined by a strong scientific focus led by Demis Hassabis, contrasting with the commercial focus of OpenAI and the engineering focus of other labs.
The organization views itself as the "engine room" of Google, managing the deployment of AI across 13 billion-user products, a scale that allows for unique A/B testing and real-world validation.
Internal dogfooding is mandatory; deep researchers and engineers are encouraged to use other models to maintain ecosystem awareness but use Gemini as their daily driver for the feedback flywheel.
The company maintains a non-zero-sum philosophy, aiming to solve human problems (like disease) rather than solely competing on benchmark scores, guided by the belief that "we can't let other people make the world a better place more than we can."
Despite the "governance" challenges of a massive corporation, the culture allows for authentic communication from leadership, with comms teams supporting rather than stifling the narrative of AI progress.