Google’s AI Coding Tools: A Manager’s Inside Guide
▼ Summary
– The median adoption date for AI coding tools was April 2024, coinciding with the release of advanced models like Claude 3 and Gemini 2.5.
– A key breakthrough enabling more capable AI coding is tool-calling, which allows models to use external functions like compiling code or running tests.
– Ryan Salva uses AI tools like Gemini CLI for both hobby coding and professional tasks, such as writing technical specification documents from under-specified issues.
– In his workflow, AI generates most of the code based on requirements and team guidelines, while he primarily uses IDEs to review the produced code.
– Salva predicts the developer role will evolve to focus more on architectural problem-solving and defining requirements, with less time spent directly writing code in an IDE.
Understanding the practical application of Google’s AI coding tools provides a clear window into the future of software development. Ryan Salva, a project manager at Google with a background at GitHub and Microsoft, offers a unique perspective from his role overseeing tools like Gemini CLI and Gemini Code Assist. His team recently published third-party research examining how developers are integrating AI into their workflows, revealing a significant shift that began around April 2024. Salva notes this period aligns with the release of advanced models like Claude 3 and Gemini 2.5, marking the arrival of what he calls “reasoning or thinking models.”
A key breakthrough during this time was the improvement in tool-calling capabilities. For coding, AI models must interact with external systems to solve problems effectively. This means an AI might need to search through files, compile code, and even run tests. The ability to use tools is what allows these models to self-correct as they progress through a task, moving beyond simple code generation to active problem-solving.
On a personal level, Salva uses a variety of AI tools for his hobby projects, predominantly through command-line interfaces. He employs a heterogeneous mix of environments, including Gemini CLI, Claude Code, and others, while using IDEs like Zed, VS Code, Cursor, and Windsurf primarily for reading and reviewing code. Professionally, his work as a product manager involves heavy documentation. He starts by using AI to transform vague GitHub issues or bug reports into detailed, outcome-driven specification documents.
The process is iterative and integrated. Salva uses Gemini CLI to build upon itself. He begins with an under-specified task, uses the AI to draft a robust requirements document, and then directs the same tool to write the code based on those specs and the team’s established guidelines. As the AI troubleshoots, it updates the requirements document, with each step creating a distinct commit and pull request for full traceability. He estimates that 70-80% of his work involves using natural language in the terminal to craft requirements and generate code, which he then meticulously reviews.
Looking ahead, Salva envisions a fundamental change in the developer’s role. For decades, the integrated development environment (IDE) has been the central hub for coding. He believes that while this will persist for some time, the balance will shift. Developers will spend increasingly more time defining requirements and less time writing raw code. This evolution may unfold over a long horizon, but it points to a future where the job resembles that of an architect. The focus will move from writing the intermediate language, the code itself, to decomposing complex problems into solvable tasks and concentrating on the bigger picture of what the software should ultimately achieve.
(Source: TechCrunch)
