Amazon published a customer case study on 29 October describing how Epsilon, a Publicis Groupe marketing-services company, offers UI testing as a managed service to its product teams. Epsilon's Agentic Platform team of 50 engineers kept its Playwright suite as the regression backbone and added Amazon Nova Act with the Amazon Bedrock AgentCore Browser tool for intent-based tests. According to Epsilon's internal benchmarks, a representative 18-step workflow takes 5-10 minutes to author as 18 natural-language instructions, against 30-90 minutes and 80-150 lines of selector-based Playwright code - a 4-6x difference in authoring effort. A workflow is explored with the model over roughly 4-5 prompt iterations until it passes; the successful trajectory is then saved and replayed deterministically, which Epsilon reports as a 100% pass rate across repeated runs. Each onboarded team runs tests in its own AWS account under a scoped role, with CloudTrail audit and S3 session recording.
Our takeRead the 100% figure carefully: it applies to replaying a recorded trajectory, not to fresh model runs. The model does authoring and exploration, and determinism comes from freezing the successful path. That is a sensible split - flakiness is paid once during the search instead of on every CI run - but coverage then depends on someone noticing a failing exploration and iterating the prompt. The open question is what happens to a saved trajectory after the UI changes: the failure mode moves from brittle selectors to a stale recording, and the same test can keep passing on an app that has moved