Google just open-sourced an ambient AI agent that debug your agents while you sleep.
It reads every failing conversation and points to the exact line of code that caused it.
Meet AQuA, the Ambient Quality Agent. 100% Open Source and built on Agent Development Kit (ADK).
Your agent returns HTTP 200, stays inside its latency budget, throws zero tool errors, and still recommends a steakhouse to a traveler whose profile says vegan.
AQuA runs beside your agent in your own Google Cloud project. It reads production traces on a schedule, after every deploy, or on demand, and it never touches the request path.
Each sweep samples up to 1,000 real sessions and grades each one for wrong tool calls, ungrounded answers, and dropped constraints.
Failures with the same root cause get grouped. A second model then rereads full transcripts and throws out any cluster the evidence doesn't support.
Clusters that pass become tracked insights, so a recurring bug shows up once with its full history instead of as a new alert every run.
Click Investigate and AQuA reads the failing traces against the exact code snapshot from that deploy. It points to the file and line range and proposes a fix, but never applies the edit itself.
The AQuA team ran it on a multi-agent travel concierge:
• It caught 6 confirmed issues in 32 test conversations and threw out 3 that didn't hold up
• The biggest one came from a single line in the agent's instructions, which let it book seats without checking if they were free
• Fixing two lines cut the seat bug from 15 conversations to 2 and the steakhouse bug from 7 to 0
• Conversations with zero issues went from 5 to 13
• The whole check cost about 12 cents per conversation
Your coding agent can pull the same insights through a CLI skill. It can apply the fix on a branch and replay the failing session before you open a PR.
You can also run the dashboard locally on a month of synthetic data without a cloud project.
Link to the GitHub repo in the comments.
It reads every failing conversation and points to the exact line of code that caused it.
Meet AQuA, the Ambient Quality Agent. 100% Open Source and built on Agent Development Kit (ADK).
Your agent returns HTTP 200, stays inside its latency budget, throws zero tool errors, and still recommends a steakhouse to a traveler whose profile says vegan.
AQuA runs beside your agent in your own Google Cloud project. It reads production traces on a schedule, after every deploy, or on demand, and it never touches the request path.
Each sweep samples up to 1,000 real sessions and grades each one for wrong tool calls, ungrounded answers, and dropped constraints.
Failures with the same root cause get grouped. A second model then rereads full transcripts and throws out any cluster the evidence doesn't support.
Clusters that pass become tracked insights, so a recurring bug shows up once with its full history instead of as a new alert every run.
Click Investigate and AQuA reads the failing traces against the exact code snapshot from that deploy. It points to the file and line range and proposes a fix, but never applies the edit itself.
The AQuA team ran it on a multi-agent travel concierge:
• It caught 6 confirmed issues in 32 test conversations and threw out 3 that didn't hold up
• The biggest one came from a single line in the agent's instructions, which let it book seats without checking if they were free
• Fixing two lines cut the seat bug from 15 conversations to 2 and the steakhouse bug from 7 to 0
• Conversations with zero issues went from 5 to 13
• The whole check cost about 12 cents per conversation
Your coding agent can pull the same insights through a CLI skill. It can apply the fix on a branch and replay the failing session before you open a PR.
You can also run the dashboard locally on a month of synthetic data without a cloud project.
Link to the GitHub repo in the comments.
As posted by Shubham Saboo on LinkedIn.