Google releases AQuA, an ambient quality agent that diagnoses production agent failures

AQuA runs unattended on a schedule, after each deployment, or on demand, reading production traces from Cloud Trace, Cloud Logging, or BigQuery. Each run samples up to 1,000 recent sessions and grades them against a nine-point checklist, then clusters the failures, verifies candidate clusters against full transcripts, and tracks them as NEW, RECURRING, or auto-RESOLVED after 14 days unseen. Raw transcripts and source snapshots stay inside the project, and AQuA never sits in the request path or writes back to the agent.

In a 32-session sweep of a multi-agent travel concierge sample, Google reports 42 structured findings that became 6 verified issues. One was planning_agent saving seat numbers to session state without checking availability with the seat-selection sub-agent. Root-cause diagnosis traced it to one line in the planning prompt, and after a one-line prompt fix on a new revision, the full-session pass count rose from 5/32 to 13/32.

Google says the review, clustering and verification cost $3.76 for the 32-session sweep, about $0.12 per session, and that on-demand root-cause diagnosis costs $0.33 to $2.47 per investigated insight. The company lists open problems including sampling sessions at scale, aligning model judges with domain experts, and replaying long trajectories when external state has changed.

If you run an agent in production, AQuA gives you a way to find failures that health checks and HTTP status codes miss, such as skipped tool calls and dropped context across sub-agent handoffs, and to tie each cluster to the revision and lines that produced it.

Its diagnoses are only as good as their citations, so check each proposed edit against the anchored lines before applying it, since the agent proposes fixes but never applies them or opens pull requests itself.