GitHub evaluates Copilot agentic harness performance

Per a June 17, 2026 GitHub engineering blog post, GitHub Copilot received two harness-level improvements to reduce token consumption in agentic sessions. GitHub VP of Product Joe Binder describes extended prompt caching -- achieving roughly 94% cache hit rates for Anthropic-backed models in VS Code -- and **deferred tool ** via a new 'tool search' mechanism that loads tool definitions on demand rather than sending every schema on each turn, cutting token overhead as tool sets grow. A second theme is Auto model selection: GitHub's internal HyDRA routing model picks the best-fit model per task using real-time health signals and task complexity. On SWE-bench, HyDRA's conservative operating point matches OpenRouter Auto at a 70.8% resolution rate while delivering 3.3x the cost savings. Per the post, Auto is already live in VS Code, github.com, and mobile, with expansion to Copilot CLI and the GitHub App planned.