The Model Eats the Scaffolding: DeepMind's Logan Kilpatrick & Tulsee Doshi on 3.5 Flash, Omni & More

2026-05-20 · Source: The Cognitive Revolution · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation, Robotics & Autonomous Systems · Depth: Intermediate, extended

Summary

Google DeepMind leaders Logan Kilpatrick and Tulsee Doshi discussed Google's AI strategy and new product launches at I/O 2026, including the Gemini 3.5 Flash model, Omni for video generation, and enhanced agent infrastructure like Antigravity and Gemini Spark. Google emphasizes cost-adjusted performance with Flash, which is three times faster and significantly cheaper than other large models, catering to consumer applications where latency and cost are critical. DeepMind now integrates models with a robust agent harness to standardize AI experiences across Google's vast product surface. The discussion also covered the practical application of recursive self-improvement, the importance of smart context usage over ever-growing windows, and Google's approach to model behavior and safety evaluations.

Key takeaway

For AI Product Managers evaluating model deployment strategies, you should prioritize models like Gemini 3.5 Flash that balance performance with cost and latency, especially for large-scale consumer applications. Consider adopting a robust agent harness approach to standardize AI experiences across your product suite, accelerating development and ensuring consistent quality. This strategy allows for rapid iteration and effective feedback loops, crucial for continuous model improvement.

Key insights

Google prioritizes cost-effective, fast AI models and integrated agentic harnesses for broad product integration and user experience.

Principles

• Emphasize cost-adjusted performance.
• Co-train models with agent harnesses.
• Standardize AI infrastructure for scale.

Method

DeepMind now provides a robust agent harness, like Antigravity, to elevate and standardize AI experiences across Google's product surface, moving beyond shipping isolated models. This co-training approach accelerates debugging, data collection, and evaluation.

In practice

• Use Gemini 3.5 Flash for cost-sensitive, low-latency applications.
• Explore agent harnesses for standardized AI product integration.
• Leverage audio input for faster code generation.

Topics

• Gemini 3.5 Flash
• AI Agent Infrastructure
• Multimodal AI
• Google DeepMind
• AI Product Strategy
• Recursive Self-Improvement

Best for: CTO, VP of Engineering/Data, AI Engineer, Director of AI/ML, AI Product Manager, AI Scientist

Related on AIssential

• Google CEO: Agents, Open Source, Race to AGI, Cybersecurity, Chips, China — Agents will fundamentally reshape internet interaction, demanding trust, control, and efficient, responsible AI development.
• 😺 Google’s sharpest brain yet? 🧠 — Google's Gemini 3.1 Pro now leads in overall intelligence and factual accuracy, offering a cost-effective, advanced AI model.
• The Agent Harness: Why a Better Model Won’t Fix Your Production Agent — Production agent failures stem from the surrounding "harness" (prompt, tools, middleware), not the underlying model.
• Gemini 3 Flash, GPT-Image-1.5, Skills vs MCPs, and Our 2025 Model Reviews - EP99.29 — Gemini 3 Flash offers surprising intelligence and efficiency, while the "Skills" paradigm promises structured, repeatable AI workflows.
• Gemini Flash Gets Pricey, AI Act Delays, Agents Drive Online Traffic — AI is creating new roles, driving up model costs, reshaping regulations, and advancing agentic capabilities and image generation.
• Google I/O 2026: Gemini 3.5 Flash, Omni, and Google’s Agent Stack — Google's strategy shifts from chatbots to agentic execution and multimodal world models, prioritizing speed and integration.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by The Cognitive Revolution.