Google released AQuA, a reference implementation that samples up to 1,000 recent agent sessions, groups the problems it finds, verifies each group and points to file and line ranges in a source snapshot from deployment. Quality criteria go in a goal.md file. In a replay of 32 travel-booking sessions, seat-selection bypasses fell from 15 sessions to two after prompt changes. These were controlled replays, not production traffic, and AQuA does not apply fixes itself.
