AI
Agents as a Core Layer of the Web
Sundar Pichai said agents will separate chores like DMV renewals from delightful web experiences. Developers already use agentic workflows for coding. He predicted agents become a core web layer but won’t replace direct browsing for entertainment or gift shopping. Users seek control and transparency as agents take on more tasks. He compared the shift to Waymo: people trust after demonstrated safety. The key is gradual capability expansion, starting with first-party services before full MCP and browser use.
Trust Through Gradual Capability Deployment
Pichai argued users already trust agents: spam filters and Waymo are examples. Google’s Gemini Spark agent begins with Gmail and Calendar before expanding to third-party MCP and full browser use. This phased approach builds trust through demonstrated safety. Users are willing to sit in a Waymo after years of data proving its safety. The company prioritizes user control and transparency as capabilities grow. Feedback loops improve the product.
Code Mender and Autonomous Security
Google announced Code Mender, an internal tool now being shared externally. It identifies vulnerabilities, generates patches, tests them, and deploys fixes automatically. It runs 24/7. Pichai said it combines with the recent Viz acquisition for real-time monitoring. The system responds to rising AI-enhanced cyber attacks. This mirrors responsible disclosure practices like Project Zero’s 90-day notification, but automated end-to-end.
Open Source Strategy and the China Question
Pichai defended Google’s balanced open-source approach, citing Chromium, Android, Kubernetes, and Gemma models. He said upfront model training costs make releasing large frontier models difficult. On Chinese open-source models, he argued companies should choose based on reliability and security, not origin. Open source with proper licensing and community oversight builds trust regardless of geography. The US must focus on staying at the frontier.
Flash Models and Cost Efficiency
Pichai emphasized flash models as critical for widespread AI adoption. 3.5 Flash is remarkably cost-efficient per token. Many CIOs worry about AI budgets blowing up. Flash models shine in agentic workflows requiring repeated calls. Google internally blends Pro and Flash models. The priority is making the best models available to billions, not just pushing the absolute frontier regardless of cost. This trade-off is deliberate: a larger model might improve the frontier but fewer can use it.
Compute Bottlenecks Across the Stack
Pichai described parallel bottlenecks in AI compute: permitting and constructing data centers, power availability, and core memory components. Solving one bottleneck reveals another. He noted costs are rising and memory prices affect planning. Google constantly makes trade-offs between model capability and accessibility. The bottlenecks are systemic across all layers of the stack. Demand exceeds supply, driving emphasis on efficient models like 3.5 Flash.
Notable Quotes
I do expect agents to be a core part of how we use the web but that doesn’t mean it’ll take away from people use the web for a lot of reasons Sundar Pichai · ▶ Watch (3:11)
Code Mender is a product which we use internally and which you’re building to share externally. Sundar Pichai · ▶ Watch (10:53)
If it is not fundamentally changing what’s out there already in terms of the state-of-the-art, I think it’s definitely okay to put it out. Sundar Pichai · ▶ Watch (12:41)
We have been pretty aggressive in deploying agent Agentic workflows. Our internal security teams use agentic workflows to help detect vulnerabilities. Sundar Pichai · ▶ Watch (9:55)
3.5 flash is remarkably costefficient Sundar Pichai · ▶ Watch (23:30)
Key Takeaways
- Pichai predicts agents become a core web layer, separating chores from browsing.
- Code Mender automates vulnerability detection, patching, and deployment 24/7.
- Google prioritizes flash models for cost efficiency in agentic workflows over pure frontier size.