Defining the agentic AI era

Key Concepts

• Agentic Era: A shift in AI development where models move beyond simple chat to autonomous, multi-step task execution.
• Full-Stack AI: Google’s integrated approach combining hardware (TPUs), model architecture (Gemini), and product application.
• Gemini 3.5 Flash: A high-performance model optimized for reasoning, coding, and agentic workflows.
• Antigravity: A framework/SDK enabling developers to build and deploy agentic applications.
• Gemini Spark: An "always-on" agent designed for asynchronous, long-running tasks.
• Amdahl’s Law (in AI): The principle that the speed of an agentic system is limited by the speed of the tools it uses, not just the model itself.
• Gemini Omni: A multimodal model capable of processing and producing diverse inputs/outputs (text, audio, video).

---

1. The Agentic Era and Model Evolution

The panel emphasized that the "Agent Era" is in full swing, driven by the transition from static models to systems capable of long-horizon reasoning.

• Gemini 3.5 Flash: Koray Kavukcuoglu noted that the focus for the 3.5 series was improving coding and agentic workflows. The model is designed to handle complex, multi-step tasks that were previously bottlenecks.
• Full-Stack Symbiosis: Jeff Dean highlighted that scaling these agents requires a tight integration between hardware (8th-generation TPUs) and software. The separation of training and inference chip designs allows for the low-latency performance necessary for agentic responsiveness.

2. Redefining Search and User Experience

Liz Reid described the integration of Gemini into Search as the "biggest upgrade to the search box in 25 years."

• Latency vs. Value: The team is evolving how they measure latency. If an agent performs a complex task (e.g., planning a multi-week trip), users are willing to wait longer than they would for a simple fact-check.
• AI Search: Search is no longer just a retrieval tool; it is an agentic experience that synthesizes data from various backends (finance, sports, local, travel) to provide a cohesive answer.

3. Agentic Workflows and Tooling

A major theme was the shift from human-speed to machine-speed software development.

• Tool Velocity: Jeff Dean explained that agents are often bottlenecked by the tools they use. By rewriting internal tools (e.g., moving from Python to Go), the team achieved 10x–20x speed improvements.
• Asynchronous Productivity: Josh Woodward and Koray Kavukcuoglu highlighted that agents like "Spark" allow for asynchronous work. Users can delegate tasks (e.g., drafting emails, researching, or managing calendars) and review the results later.
• Machine-Readable Context: Teams are moving away from traditional Product Requirement Documents (PRDs) toward machine-readable formats like DESIGN.md files, which allow models to autonomously understand and implement design languages.

4. The Future of Interfaces

The panel discussed how interfaces will evolve as agents become more capable.

• Voice-Native Interfaces: Josh Woodward noted that the team is exploring "voice-native" experiences that go beyond simple commands, potentially moving toward "mission control" style dashboards.
• Personalization: Liz Reid suggested that the interface should adapt to the user’s specific way of thinking and processing information, rather than being a one-size-fits-all dashboard.

5. Notable Quotes

• Jeff Dean: "If you make the model infinitely fast, Amdahl's law says if you're spending half your time in tools, you're not going to get anything better than 2x speedup."
• Liz Reid: "People's willingness to wait is in part based on how much work you are taking off."
• Koray Kavukcuoglu: "We are in that in-between of not academic research anymore... It is a technology that is going out in the world."

6. Synthesis and Conclusion

The transition to the Agentic Era represents a fundamental change in how software is built and used. By leveraging "full-stack" capabilities—from custom TPU hardware to advanced models like Gemini 3.5—Google is enabling a shift where agents act as autonomous collaborators. The primary takeaways are:

1. Efficiency: The future of productivity lies in asynchronous delegation to agents.
2. Customization: AI allows for the creation of "bespoke software," where users can generate tools tailored to their specific needs on the fly.
3. Infrastructure Transformation: To fully realize the potential of agents, the underlying software infrastructure must be optimized for machine-to-machine communication rather than just human-to-machine interaction.