The future of software development

Key Concepts

• Gemini 3.5 Flash: Google’s latest, most capable model, optimized for high performance, speed, and complex tool use.
• Agentic Workflows: The shift from simple chat-based interactions to long-running, multi-step processes where AI acts as an autonomous agent.
• Antigravity: A Google internal/external platform serving as a conduit for the Gemini ecosystem, enabling complex agentic tasks.
• Product Flywheel: The symbiotic relationship between model research, product development (harnesses), and user feedback loops.
• Vibe Coding vs. Agentic Engineering: The distinction between rapid, creative prototyping ("vibe coding") and building robust, production-grade systems ("agentic engineering").
• Information Retrieval (IR) Bottleneck: The challenge of providing models with accurate, real-time, and secure data to perform tasks without hallucination.
• Multimodal Interaction: The ability of models to process diverse inputs (video, audio, text) to perform complex tasks like choreography or system monitoring.

---

1. Model Capabilities and Performance

• Gemini 3.5 Flash: Tulsee Doshi highlights that this model outperforms 3.1 Pro on most benchmarks. Its primary strength lies in tool use, long-running coding tasks, and productivity applications.
• Speed and Intelligence: Varun Mohan notes that Flash achieves over 200 tokens per second. Despite the speed, the model is now capable of "thinking" for 30–40 seconds to solve complex problems, such as building an entire operating system (requiring ~15,000 model invocations).
• Dynamic FPS: The Gemini API now supports dynamic frames-per-second (FPS) adjustments, allowing for high-fidelity analysis of fast-moving visual data (e.g., sports or dance).

2. The Evolution of Software Development

• From Chat to Agents: Michael Gerstenhaber explains that the "Gemini Enterprise Agent Platform" (formerly Vertex AI) reflects a shift from simple chat UIs to agentic workflows.
• Interaction Patterns:
  • Intelligence-bound: Tasks requiring deep reasoning (e.g., writing complex code).
  • Latency-bound: Tasks requiring immediate responses (e.g., customer service policy application within 250ms).
  • Scale-bound: Tasks requiring infinite scale (e.g., internet-wide content moderation).
• The "Coworker" Metaphor: Models and their harnesses should be treated as coworkers. Developers should build systems that "fail" today because the model isn't smart enough yet, rather than failing because the system architecture is flawed.

3. Methodologies and Frameworks

• The Product Flywheel: Varun Mohan emphasizes that research teams must use the same "harnesses" (the infrastructure surrounding the model) that external users do. This creates a feedback loop where researchers experience the same struggles as developers, preventing the "benchmark-only" trap.
• Agentic Engineering: Unlike "vibe coding," this requires:
  • Asynchronous Tool Calls: Allowing agents to run long-running jobs (e.g., training models) without blocking the main interaction.
  • Subagent Delegation: The ability for a primary agent to select and delegate tasks to specialized sub-models.
  • Robust Testing: Using agents to generate their own integration tests (e.g., using Playwright) to ensure new features don't break existing functionality.

4. Key Arguments and Perspectives

• The Bottleneck Shift: As coding becomes easier, the bottleneck shifts to "Taste" (identifying real user pain points) and "Information Retrieval" (accessing secure, proprietary data).
• Security and Credentialing: Michael Gerstenhaber highlights that banks are now using agents for "Know Your Customer" (KYC) processes. The barrier was never model intelligence, but rather the software engineering challenge of credentialing agents to access sensitive data without exfiltration.
• Interface Design: The panel agrees that the future of interaction is not just text. Voice, gesture, and visual pointing are becoming critical. The agent should choose the interface (phone call, email, or chat) based on the urgency and context of the task.

5. Notable Quotes

• Michael Gerstenhaber: "It’s not a spectrum of intelligence. It’s not a spectrum of latency. But it’s actually an easy to describe, intentional reason to use a lot of the models."
• Tulsee Doshi: "When building is so easy, it’s very easy to then build for problems that maybe even aren’t real problems."
• Varun Mohan: "The people that are going to shine are the people that are going to be incredibly high agency. They can make hard calls."

6. Synthesis and Conclusion

The transition to an agentic era requires a fundamental shift in how we build software. The "model" is no longer just a text generator; it is a reasoning engine that must be integrated into a robust "harness" capable of handling long-running tasks, asynchronous tool calls, and secure information retrieval. The most successful developers will be those who focus on high-agency decision-making—identifying the right problems to solve—and building systems that leverage the model's ability to delegate tasks, test itself, and interact through diverse, human-centric interfaces.