Content by Julia Kasper (2)

Julia Kasper (VS Code Eval Team) breaks down what an ultra-simple “write HELLO.txt” agent eval revealed after 50,000+ runs: models vary widely in tool-call discipline, planning overhead, and output-token cost, and those differences matter for latency, billing, and automatic model selection in VS Code.

Julia Kasper explains the “coding harness” that powers GitHub Copilot’s agent experience in VS Code: how prompts and workspace context are assembled, how tools are exposed and executed, and how the agent loop is controlled and evaluated as models and providers change.