Release cadence is now measured in hours. A single business capability spans dozens of microservices, several APIs, cloud infrastructure that changes underneath the application, and machine-learning components whose behavior is statistical rather than specified. The surface a test team is asked to cover grows faster than any team can staff against it.

Eight studies work through the questions that cost the most in a modern delivery pipeline: which tests to run first, which inputs to generate, where the next failure is most likely to hide, and which automation should repair itself instead of raising a ticket. Graph neural networks order a regression suite and explain the order. Bayesian inverse reinforcement learning recovers the objectives experts actually optimize. A language model paired with a symbolic engine generates tests that are both plausible and reachable. Causal discovery repairs broken UI locators on evidence.

Each chapter reports the operating cost beside the gain: inference latency in a build gate, retraining under drift, and the risk of an autonomous agent acting on a live system.