How Prompt Tuning Improved GPT-5.5 in VS Code

The VS Code team ran a two-week A/B experiment with OpenAI to tune the GPT-5.5 system prompt, testing whether nudging the agent to explore less and validate sooner would cut costs and latency without hurting quality. Two prompt variants were tested against a control on live traffic. The winning variant restructured the agent loop into explicit before/after-edit sections, delivering 5.7% faster median edits, 8.5% fewer tool calls, and significant token savings at the upper tail.