Google's Dev Tools Lead on Integrating AI into Software Development

As Google’s project manager for developer tools, Ryan Salva has a unique vantage point on how AI tools are transforming software development. With a background at GitHub and Microsoft, he now leads initiatives like Gemini CLI and Gemini Code Assist, guiding developers into the evolving landscape of agentic programming.

His team recently published independent research examining how developers are actually using AI tools and what improvements are still needed. I spoke with Salva about the report and his personal experiences with AI-powered coding assistants.

This interview has been edited for length and clarity.

Google conducts an annual developer trends survey, but this year's report places a strong emphasis on AI tools, particularly how willing developers are to adopt agentic approaches in their workflows. Did any findings in the research surprise you?

One particularly striking discovery was the median date when developers began integrating AI tools into their work: April 2024. This aligns closely with the releases of Claude 3 and Gemini 2.5, marking the emergence of reasoning-focused models. Around the same period, we saw significant advancements in tool-calling capabilities.

For coding tasks, models must effectively leverage external information to solve problems—whether that involves searching through code, compiling it, or running unit and integration tests. The breakthrough in tool-calling is crucial because it enables models to self-correct as they progress through complex tasks.

How do you personally use AI coding tools in your own work?

Join 10k+ tech and VC leaders for growth and connections at Disrupt 2025

Netflix, Box, a16z, ElevenLabs, Wayve, Sequoia Capital, Elad Gil — just a few of the 250+ industry leaders hosting 200+ sessions packed with insights to drive startup growth and sharpen your competitive edge. Don’t miss the 20th anniversary of TechCrunch and this opportunity to learn from top tech innovators. Book your ticket by September 26 to save up to $668.

Join 10k+ tech and VC leaders for growth and connections at Disrupt 2025

Netflix, Box, a16z, ElevenLabs, Wayve, Sequoia Capital, Elad Gil — just a few of the 250+ industry leaders hosting 200+ sessions packed with insights to drive startup growth and sharpen your competitive edge. Don’t miss the 20th anniversary of TechCrunch and this opportunity to learn from top tech innovators. Book your ticket by September 26 to save up to $668.

Most of my coding now is for personal projects, and I primarily work with command-line tools like Gemini CLI. I also experiment with Claude Code and Codex. Since terminal-based tools are rarely used in isolation, I regularly switch between different IDEs—Zed, VS Code, Cursor, and Windsurf—to stay current with industry developments and understand how various tools are evolving.

Professionally, product managers spend a lot of time working with documents, so my first step is often using AI to help draft specifications and requirements documents.

I'm curious about the practical side. You're using Gemini CLI to help build Gemini CLI itself, but I assume the process isn't entirely automated.

A typical development task starts as an issue—often a GitHub issue reporting a bug. To be honest, these are frequently under-specified. I use Gemini CLI to generate a more thorough requirements document in Markdown, typically producing about 100 lines of technical, outcome-driven specifications. Then, I use Gemini CLI to write the actual code based on those specs and our team’s established guidelines.

Our engineering team maintains several layers of rules and Markdown documentation that inform the model about our workflow—covering testing protocols, dependency management, and other standards. When generating code, the model references these documents to ensure consistency.

As Gemini CLI handles troubleshooting, I have it update the requirements document to track progress: "Fixed this step, moving to the next," and so on. Each update becomes a separate commit and pull request, making it easy to review or revert changes as needed.

I'd estimate that 70-80% of my work involves using natural language in the terminal with Gemini CLI—first to refine requirements, then to generate the majority of the code. I review the output in whichever IDE I'm using, but increasingly, I rely on the IDE more for reading code than for writing it.

Do you think traditional code writing will persist, or will we shift entirely to terminal-based workflows?

For the past thirty years, the IDE has been the central hub for software development, complemented by the browser and terminal.

I believe that structure will remain relevant for some time, but I expect we'll gradually spend more time defining requirements, with less time dedicated to hands-on coding in the IDE. This transition will likely unfold over an extended period.

Naturally, there's concern about what this means for the future of software development careers. If we're not directly writing code in a decade, what does that mean for developers? Will their roles still exist?

I see the developer's role evolving to resemble that of an architect. Your focus will shift to decomposing complex problems into manageable tasks and considering the broader vision of what you're building, rather than getting bogged down in the specifics of programming languages and machine code.