Improving token efficiency for GitHub Copilot in VS Code

GitHub Copilot's agentic harness in VS Code now reuses cached prompt prefixes to cut token costs and latency. Cached tokens can be up to 10 times cheaper. A/B experiments confirm task success rates hold or improve while token usage drops through a steady stream of small efficiency gains.