Copilot's app has a slash command whose entire job is to disagree with you.

GitHub documented the surface August 6: /spar pressure-tests your approach by questioning assumptions and naming risks, /rubber-duck runs an independent review using a different model to surface blind spots. The rest map to modes: /plan, /autopilot, /orchestrate for parallel work across multiple repos, /create-canvas. Adversarial review as a first-class command instead of a prompt convention everyone reinvents is the design choice worth stealing. Separately, the usage metrics API now reports agent app activity, which is the change that unblocks measuring agent rollout at all.