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

Summary

In this episode, Logan Kilpatrick frames Google’s AI strategy around a coming shift from standalone models to agentic harnesses layered on top of them—and then eventually inside them. He says the current wave of harnesses, tools, containers, search, and code execution may have only about a 12-month shelf life before models digest much of that scaffolding. He also emphasizes that Google’s rollout pace is intentionally cautious at Search scale, with product success defined by customer outcomes rather than eyeball time. On the technical side, he highlights coding as the strongest current agent use case, describes 3.5 Flash as a coding leap driven by post-training, and positions Omni as a single multimodal model meant to replace a sprawl of separate text, audio, image, music, and video systems.

Notes

• Google is shifting the core AI through-line from the Gemini model/API to the agent harness layer.
• Coding is treated as a specialized but also general-purpose agent harness.
• Google’s agentic products are still mostly at a crawl, not a run.
• Google’s scale forces caution, especially on Search.
• The Gemini app is closer to walk than most other Google surfaces.
• Agentic AI is still early enough that there is a long tail of product experimentation.
• Google wants AI products to maximize customer outcomes, not attention.
• Search and AI can be positive-sum rather than cannibalistic.
• Long-running agent tasks are becoming a more relevant measurement target.
• Long-running agents and coding agents are both top priorities for DeepMind.
• Anti Gravity is part product, part internal harness, and part developer ecosystem.
• 3.5 Flash is positioned as a coding model stronger than prior Pro models.
• The coding gain in 3.5 Flash came from post-training, not pre-training.
• Google and DeepMind have a large internal feedback flywheel from 100,000-plus engineers.
• He describes coding as already feeling like narrow superintelligence.
• He expects the earliest vertical superintelligence gains in highly verifiable domains.
• Video games are bottlenecked more by scaffolding than raw model quality.
• Early AI Studio data showed games as a major app category.
• He says finance-related apps were also roughly 20% of usage.
• Vibe-coded games may be possible this year if scaffolding improves.
• Short-term game generation likely needs coding agents plus game engines.
• Omni is being framed as a single model, not a router across separate media models.
• Omni Flash is the first iteration and later versions should be more capable.
• Omni’s generative value is that it can change the scene without changing the person.
• AI Studio is designed to surface more of Google’s ecosystem through one workflow.
• He built his first Android app in AI Studio.
• AI Studio had already produced 350,000 Android apps in a week.
• Many of those apps are personal tools that likely would not have existed otherwise.
• He argues operating systems still matter because they unlock richer native experiences than web chat.
• Modern models are no longer just weights; they are sprawling systems around the weights.
• That modern stack includes tools, search, code execution, containers, and agent harnesses.
• He predicts much of today’s harness layer will be upstreamed into the model within about 12 months.
• External scaffolding will still matter for search and code execution.
• He proposes a benchmark for how well models adapt to different harnesses.
• Startups can still win by focusing tightly and moving faster than labs.
• Coding tools help startups close the gap with incumbents.
• DeepMind is described as scientific, mission-driven, and portfolio-oriented.
• He says the original DeepMind culture still looks like “get smart people together and solve the problem.”
• He ties DeepMind’s mission to real human problems like disease, not leaderboard chasing.
• DeepMind collaborates across Android, Cloud, Gmail, and Workspace.
• Deploying Gemini across billion-user products is a rare systems problem at Google.
• Google’s comms and marketing teams are portrayed as supportive partners.