Google gives out AQuA, an open source tool for silent agent faults

Google released AQuA, short for Ambient Quality Agent, an open source tool that reads the logs of AI agents and finds silent faults. A silent fault is an agent that ignores a user preference but gives a good result. In a Google replay test, sessions that did not do a check of seat availability went from 15 to 2. The test did not use live data.

How it works

AQuA operates adjacent to an agent in Google Cloud. It reads up to 1,000 sessions each time. It compares sessions that have a fault with a copy of the code from deployment. It gives the part of the code that is the cause.

The numbers

Google did a test of 32 sessions of a travel app. A verifier rejected 3 groups that had no data and kept 6 problems. After a change to the instructions, sessions with no fault in all checks went from 5 to 13.

What is not known

Google tells that a sample of sessions can not find a fault that only some sessions show. A replay can copy only part of the condition of external systems. Dmytro Aprelenko of Redcore tells that the group checks decrease incorrect alarms. They do not make sure that all sessions show the same fault.

Insight: Agent monitors find faults but do not stop agents

AQuA makes no repair. Agent Anomaly Detection of Google found 87% of dangerous steps but only 4 of 18 poisoned memory examples. Monitors read logs after the work is done, and the person who gives approval for a change is the control that works. AQuA reads Google Cloud logs, and this reads as a move to pull agent checks to that cloud. Select who gives approval for each agent change.