How Google’s dev tools manager makes AI coding work (techcrunch.com)

Google’s developer-tools project manager Ryan Salva walked through how his team’s new third‑party research and his own work practices reflect a fast shift toward “agentic” AI coding. The report found the median date developers began regularly using AI tools was April 2024 — coinciding with launches like Claude 3 and Gemini 2.5 — marking a turning point where reasoning models plus robust tool‑calling (grep, compile, run unit/integration tests) let models self‑correct. Salva uses Gemini CLI and Gemini Code Assist alongside multiple IDEs (Zed, VS Code, Cursor) and other models, and estimates he spends 70–80% of his hands‑on workflow in a terminal driving natural‑language prompts to craft specs, generate code, and let the agent produce commits and pull requests that can be reviewed or rolled back. Technically, the workflow emphasizes structured requirements documents and team “way of working” Markdown (testing, dependency rules) fed to the model so generated code adheres to norms, and tool-calling for external information and runtime feedback. The significance is twofold: tool integration and agentic workflows materially change how code is produced and verified, and they shift developer roles from line‑level implementation toward higher‑level architecture, decomposition, and specification. That signals major product, process and education changes for the AI/ML and software communities as agents mature.