AQuA
by Google · Agent · Text & code
unproven
Reference tool for finding agent failures; results come only from Google's small demo
Google open-sourced AQuA, a reference implementation that analyzes production agent sessions and investigates root causes against deployed source revisions
AccessOpen codeLicenseopen source
Why we say so
Google reports 42 findings across 32 travel-app sessions, six verified, and passing sessions rising from 5 to 13 after prompt changes, all in controlled replay rather than production. InfoQ describes the design and notes it proposes fixes but does not apply them. No outside team has reported using it, and Google itself says sampling can miss rare failures.
Evidence
- Nov 13Google lists limits: sampling can miss rare failures and replay may not reproduce external state; results came from controlled replay, not production
- Nov 13On 32 travel-app sessions it found 42 findings in nine clusters; three were rejected, leaving six verified issues. After prompt changes and replay, passing sessions rose from 5 to 13