Google has open-sourced AQuA, an “ambient quality agent” that looks for behavioral failures hidden behind technically successful AI-agent responses. Examples include ignoring a user preference or skipping a required tool call even though the service returned without an infrastructure error.
AQuA reads telemetry from Cloud Trace, Cloud Logging or BigQuery outside the live request path. Each run samples up to 1,000 recent sessions, reviews them against configurable criteria, clusters similar findings and verifies candidate patterns. For repository-related failures, it compares traces with an immutable snapshot of the deployed source and cites relevant files and lines.
In a controlled replay of 32 sessions from a travel application, the system found 42 potential issues in nine clusters. Verification rejected three clusters. After prompt changes, seat-check bypasses fell from 15 sessions to two, dietary-preference omissions dropped from seven to zero, and sessions passing every check rose from five to 13.
Those numbers came from replay rather than live production and do not establish performance across other agents. Sampling may miss rare failures, and a replay cannot reproduce every external system state. AQuA deliberately stops at diagnosis: it may propose a change, but it does not edit code or open pull requests, leaving the remediation decision with developers.