Reliable agents need an execution harness, a way to judge outcomes, and a way to learn from failures after release. This five-part series connects those responsibilities without giving a quality observer unrestricted production access. Start with harness contracts, then examine a real observer, build a small offline example, and finish with evaluation and deployment gates.
AQuA is Google's Ambient Quality Agent, not a Meta product. See the Google announcement for its origin and purpose. The Google implementation is an alpha reference implementation; our source review is pinned to its v0.1.0 release. The articles distinguish documented implementation behavior from Sandeep Belgavi's original teaching examples and proposed operating practices.
Read the architecture articles before the example. The example uses synthetic records and Python's standard library; it does not install Google AQuA, contact a model, or demonstrate a cloud deployment. The final article explains how to judge findings and decide whether an observer is ready for a controlled production trial.
Continue with the existing Harness Engineering collection for execution-system design, or explore the labs collection for interactive practice. These articles belong together in Evals because observation becomes useful only when its findings can be checked.
The five-guide reading path
Harness engineering: make agent outcomes inspectable
Design an execution harness whose permissions, tool effects, checkpoints, and completion claims can be evaluated independently. A concrete shipment example connects runtime contracts to an ambient quality loop.
Guide 02Google AQuA: what it is, where it lives, and how to try it
Locate Google's Ambient Quality Agent, understand its alpha status and supported setup paths, and choose an honest first experiment before connecting production telemetry.
Guide 03Inside AQuA: architecture, evidence, storage, and verification
Read Google's documented quality workflow alongside pinned implementation details, including a crucial verification default. Then design explicit boundaries for evidence, storage, and human-controlled repairs.
Guide 04Build an offline ambient quality worker with Python
Run a small standard-library example with synthetic tool and outcome events, duplicates, and injected text. Learn deterministic checks and their limits before adding models or cloud services.
Guide 05Evaluate ambient quality agents before trusting their findings
Measure coverage, precision, recall, evidence quality, and repair effectiveness separately. An original worked example and staged deployment plan show how to build observability without confusing silence with success.