AWS: Epsilon runs plain-English UI tests on Nova Act

AWS published a customer write-up on 29 October 2026 describing how Epsilon, a data and marketing-services company in Publicis Groupe, delivers UI testing as a managed platform service. The AWS authors say Epsilon's Agentic Platform team, 50 engineers who build shared agent infrastructure, uses Amazon Nova Act and the Amazon Bedrock AgentCore Browser Tool. The post is vendor-published and promotes both products.

Added to Playwright, not instead of it

The post says Epsilon's Playwright regression suite was not broken and remains its backbone. The gap it names is that scripted tests are tied to selectors and page structure, so cosmetic changes can break them even when the feature works. Epsilon layered natural-language tests on top, such as 'Confirm the rewards balance is visible', rather than rewriting what already worked.

How the service is built

A product team creates a workflow through a platform API with a target URL and a list of plain-English steps. A Celery background worker opens a managed browser session, runs each step through Nova Act over a Chrome DevTools Protocol connection, and records a screenshot and a pass or fail per step. Results, screenshots and HTML reports go to Amazon S3.

The capability is exposed as one agent with two skills, a recorder that explores a scenario and writes the steps, and an executor that runs them. Each onboarded team gets a cross-account IAM role, so its tests run in its own AWS account. The post says AgentCore Browser Tool adds live view, session recording to S3 and AWS CloudTrail audit without extra configuration.

The reported numbers

• Authoring: about 5 to 10 minutes for a representative 18-step workflow, against 30 to 90 minutes for an equivalent Playwright script, which the post estimates at roughly 80 to 150 lines. Epsilon's internal benchmarks, per AWS.
• Replay: after a workflow passes its explore phase, typically 4 to 5 prompt iterations, Epsilon saves the successful trajectory and replays it. The post reports a 100 percent pass rate on repeated replays.
• Cost: replaying a saved trajectory skips the model's reasoning, which the post says lowers latency and cost. It gives no dollar figure.

Lessons the post lists

• Write steps as intent rather than mechanics, so product managers and QA leads can review them.
• Run Nova Act in its own service with pinned dependencies. Epsilon found that the Nova Act signing endpoint targets us-east-1 whatever region the browser session is in.
• Use draft-state retries to find a working path, then replay the saved trajectory in active state.
• Account for the browser's default viewport when a step says 'scroll down until the field is visible'.

What the post leaves out

The 100 percent pass rate describes replays of paths that had already passed, not how often a fresh natural-language run reaches the right outcome. The post does not say how often an agent-run test passed when the app was actually broken, or whether it passed on a wrong page. It gives no cost per run, no count of teams using the service beyond 'multiple', and no baseline for maintenance effort over time. The authoring comparison is an estimate for one workflow.

For a team weighing the same approach, the portable pattern is to keep deterministic tests for regression, use language-driven tests for flows that churn, and record a successful path before relying on repeat runs. Test the second step by breaking the app on purpose and checking that the agent reports a failure.