VS Code ships weekly via AI agents

Scaling coding models is no longer just a model training challenge, but an engineering battle of offline evaluation pipelines and automated agent triage to support rapid release loops.

* Custom offline evaluation frameworks like VSC-Bench are critical to verifying that model updates improve performance in real-world scenarios rather than just static benchmarks.

* The transition from monthly to weekly releases requires replacing manual review bottlenecks with AI-agentic automated triage, PR generation, and UI screenshot comparisons.

* Developer feedback loops and code survival metrics (the percentage of AI code actually committed) drive the fine-tuning process, optimizing prompt construction directly in the editor codebase.