how google’s dev tools manager makes ai coding work

The Evolving Landscape of Coding with AI: A Conversation with Ryan Salva

Ryan J. Salva, Google’s senior director of product management for developer tools, possesses unique insight into the transformative impact of artificial intelligence on the coding process. Having previously held positions at GitHub and Microsoft, he currently oversees the development of tools such as Gemini CLI and Gemini Code Assist, guiding developers toward the emerging paradigm of agentic programming.

New Research on AI Tool Adoption

Salva’s team recently published new research on Tuesday detailing how developers are integrating AI tools into their workflows – and highlighting areas where further advancements are needed. A discussion with Salva explored the report’s findings and his personal experiences utilizing AI-powered coding assistants.

The following interview has been condensed for brevity and clarity.

Developer Adoption Timeline

Google conducts an annual survey of developer trends, but this year’s report particularly emphasizes AI tools and the willingness of developers to embrace an agentic approach to programming. Were there any findings in the research that particularly stood out to you?

A notably interesting discovery was the median timeframe for when developers began experimenting with AI tools. The data indicated April 2024, coinciding with the release of Claude 3 and Gemini 2.5. This period marks the arrival of reasoning and thinking models, and concurrently, significant improvements in tool-calling capabilities.

Effective coding necessitates the ability to access and utilize external information for problem-solving. This may involve searching through files, compiling code, running unit tests, and executing integration tests. The advancement of tool-calling is crucial, enabling models to self-correct and refine their solutions iteratively.

Personal Use of AI Coding Tools

Could you describe how you personally utilize AI coding tools in your work?

My coding endeavors are currently focused on personal projects, and I primarily rely on command-line based tools. This includes Gemini CLI, alongside some Claude Code and Codex. I don’t typically use these tools in isolation, and I experiment with a variety of Integrated Development Environments (IDEs). I utilize Zed, VS Code, Cursor, and Windsurf, driven by a desire to understand the evolving industry landscape.

Professionally, product managers frequently work within documentation, so I initially leverage AI to assist in drafting specifications and requirements documents.

Building with AI: Gemini CLI

I’m interested in the process of building Gemini CLI with Gemini CLI. It seems unlikely that it operates entirely autonomously.

A development task often originates as an issue, potentially a bug report submitted via GitHub. Frequently, these issues are initially vaguely defined. I then employ Gemini CLI to create a more comprehensive and detailed requirements document in Markdown format. This typically results in approximately 100 lines of technical, yet outcome-focused, specifications.

I then instruct Gemini CLI to generate code based on this specification, adhering to the established team guidelines documented in various Markdown files. These documents outline our processes for testing, dependency management, and other key aspects of our workflow. Consequently, the generated code aligns with our established standards.

As Gemini CLI progresses through troubleshooting, it updates the requirements document, indicating completed steps and outlining the next tasks. Each update is committed as a separate pull request, allowing for easy review and rollback if necessary.

I estimate that 70% to 80% of my work involves interacting with the terminal using natural language, utilizing Gemini CLI to refine requirements, and allowing it to generate the majority of the code. I then review and analyze the code within an IDE, primarily using it as a reading environment rather than a writing one.

The Future of Code

Do you foresee a future where traditional computer code becomes obsolete, or will we increasingly rely on terminal-based interactions?

For the past three decades, the IDE has been the central hub for software development, alongside the browser and the terminal window.

While this remains largely true, I anticipate that we will dedicate more time to defining requirements, and the time spent directly within the IDE will gradually decrease. However, I believe this transition will unfold over a considerable timeframe.

Implications for Software Developers

There is considerable concern regarding the implications of these changes for the future of software development. If, in ten years, developers are no longer primarily focused on code, what will their role be? Will there still be a demand for their skills?

The role of a developer will evolve to resemble that of an architect. It will involve tackling complex problems and breaking them down into smaller, manageable tasks. Developers will need to focus on the overall vision and desired outcomes, rather than the specific language used to translate those ideas into machine code.