Verified, then logged.
Sun 22 Nov 2026 · 00:20 KST / 15:20 UTC한국어 | ENSearch /

02 / IT · Analysis

Google releases AQuA to diagnose AI agent failures

AQuA links recurring failures to deployed code, with session replay used to check prompt changes.

A hand holds a smartphone over a wooden table, its home screen nearly empty except for a few plain app icons, as the thumb drags one away to delete it. AI-generated image

Google has released AQuA as open source, according to InfoQ. The tool is aimed at developers investigating problems in AI agents running in production.

InfoQ reports that AQuA checks recurring failure patterns and investigates their causes using source code snapshots from deployment. This connects failure records with the code used during execution.

The report also describes developers replaying sessions after changing prompts. That process connects failure diagnosis with checks on a subsequent fix.

Identifying recurring failures

InfoQ describes AQuA as a reference implementation for analysing production agent sessions. A reference implementation shows how a system's functions fit together and operate.

According to the report, its functions cover both checking recurring failures and investigating causes against deployed code. The investigation therefore extends beyond what happened in a session to the source snapshot behind it.

Replaying sessions after prompt changes

InfoQ's example covers agents bypassing seat selection and omitting dietary preferences. Developers changed the prompts and then replayed the sessions to check the results.

In the process InfoQ describes, identifying a failure pattern leads to a change and another run. Diagnosis provides the starting point for a fix, while replay provides a way to check what happens after that change.