# How VS Code Went from Monthly to Weekly Releases with AI — Harald Kirschner

## Executive summary

Harald Kirschner details how the VS Code team transitioned from monthly to weekly releases by fundamentally rebuilding their software development system around AI agents. The focus is not merely on using AI to generate code, but on establishing robust feedback loops, improving CI/CD performance, and implementing quality gates (like automated code review and staged rollouts) to maintain stability at high velocity. Key systemic improvements include achieving a 10x build speed increase with TypeScript Go, automating issue triage from 51 billion daily telemetry events, and making the codebase 'agent-ready' through skills and documentation.

## Key takeaways

- Systemic Change is Key to Velocity: Achieving high velocity requires evolving the entire software shipping system, not just increasing AI usage. The goal is to make the process of shipping high-quality code faster and more reliably.
- Code Survival Metrics Improved: The percentage of agent-written code that successfully gets committed (Code Survival) increased from 55% (using GBD 4.1) to 86% (using Cloud Opus 4.6), demonstrating growing developer trust in AI-generated code.
- Automated Quality Gates: The team implemented mandatory AI code review, automated issue triage, and error stack analysis. This process filters 51 billion daily telemetry events into actionable issues and auto-created PRs.
- Adopting Stage Rollouts: To derisk the process, VS Code moved away from 'YOLO' (release day 100% rollout) to staged rollouts, monitoring error logs and issues at each step.

## Technical details

- Codebase Preparation: Making the codebase 'agent-ready' involves creating lightweight agents and comprehensive documentation (e.g., using `agents.mmd`) to map the codebase for AI consumption.
- CI/CD Performance: Switching to TypeScript Go resulted in a massive 10x improvement in build times, which is critical when running numerous automated PRs and CI/CD loops.
- Testing and Feedback Loops: The team utilizes the Component Browser (an automated build that takes screenshots of every component change) and Playwright (for web app testing) to create tight, self-correcting feedback loops, especially for UI development.
- Issue Management: AI now handles filtering, enriching, and translating issues, assigning area owners. A human layer remains critical for reviewing agent-generated fixes and preventing duplicated issues.

## Practical implications

- Focus on building robust feedback loops (e.g., using Playwright or automated component testing) rather than just generating code.
- Treat CI/CD performance as a critical bottleneck, especially when scaling automated agent activity.
- Implement structured processes (like 'skills' or standardized documentation) to encode expert knowledge and unblock development.
- Adopt staged rollouts and continuous monitoring of telemetry data to derisk high-velocity releases.

## Topics

Software Development Lifecycle (SDLC), AI Agents, Continuous Integration (CI), Code Quality Assurance, Product Management, What 50,000 Runs of a 5-Line Eval Taught Us, Improving token efficiency in GitHub Copilot

Source: https://www.youtube.com/watch?v=I2LL_wd89-A