Vibitely Daily

Saturday, 14 November 2026, story 6 of 7

Google open-sources AQuA, a tool to trace production agent failures

The reference implementation clusters recurring failures and links them to deployed code. Google's 32-session replay showed fewer failures after prompt changes.

Illustration: A magnifying glass of translucent glass hovers over a tangle of colored threads, with a few threads of the same color pulled out and bundled neatly to one side.

What happened

Google has open-sourced AQuA (Ambient Quality Agent), a reference implementation that analyzes production AI agent sessions, verifies recurring failure patterns and investigates root causes against deployed source revisions, InfoQ reports. It runs alongside an agent in Google Cloud, outside the request path, and reads Cloud Trace, Cloud Logging or BigQuery. Each run samples up to 1,000 recent sessions.

In Google's evaluation of 32 sessions from a multi-agent travel application, AQuA found 42 findings in nine candidate clusters. The verifier rejected three, leaving six. After prompt changes and a replay, seat-selection bypasses fell from 15 sessions to two, and sessions passing all checks rose from five to 13. These were controlled replays, not production traffic.

What it means

AQuA targets failures that return a successful response while the agent ignores a preference or skips a tool call. Ordinary infrastructure monitoring can miss those. Teams running agents on Google Cloud telemetry get a template for clustering such failures and tying them to a specific code or instruction revision. The tool proposes changes but does not apply them or open pull requests, so a developer still reviews every fix.

The evidence is thin. It is one demo application, 32 sessions and a replay, and Google itself says sampling can miss rare failures and replay may not reproduce external state. The improvement figures show the loop can work in a controlled setting, not that it holds under production load. Teams on other stacks would have to adapt the telemetry layer. Watch for results from real deployments.

What to do: If you run agents on Google Cloud, try AQuA on a sample of sessions and treat its proposed fixes as leads to review, not as proof.

Source

Filed under Open source. Published .

"What it means" is analysis, not reporting. Found an error? Write to daily@vibitely.com.