Epsilon is using Amazon Nova Act to test website workflows with plain-English instructions, adding an AI-driven layer to its existing Playwright tests. The practical shift is that teams can check what a user is trying to do, rather than relying only on selectors that may break when a page’s appearance changes.
AWS AI Watch analysis
What happened
AWS describes Epsilon’s system as a managed testing service built with Nova Act and Amazon Bedrock AgentCore Browser. Teams submit workflows through a shared platform; a background runner opens an isolated browser session, carries out the steps and records results, including screenshots and pass or fail status.
A workflow can include actions such as signing in, creating an agent and checking that it appears in a list. Conditional instructions can handle variations, such as skipping an onboarding prompt if it appears. Epsilon’s example uses 18 natural-language steps, compared with roughly 80 to 150 lines of equivalent Playwright code, AWS says.
The approach is designed to complement, not replace, Epsilon’s existing Playwright regression suite. AWS says the system is operating across multiple teams. Its account of Epsilon’s implementation also describes isolated browser sessions, cross-account access controls and recorded test results.
Why it matters
A cosmetic redesign can break a test tied to a particular page selector even when the feature still works. A test framed around intent, such as checking that a rewards balance is visible, may be less brittle and easier for non-engineers to review.
The other useful change is operational: Epsilon’s platform team provides the browser and identity plumbing as a shared service, rather than asking every product team to maintain its own testing infrastructure. That makes this more than an AI agent clicking through a demo. It is a concrete example of how teams might fit agent-driven testing into existing software processes.
Our read
The strongest idea is the pairing of plain-English steps with an established regression suite. Let the agent explore the parts that resist rigid scripts, but keep deterministic tests where they already do the job. AI need not be handed the whole testing department just because it can find a button.
AWS presents Epsilon’s system as an operational example, not a controlled comparison proving that Nova Act is more reliable or cheaper than Playwright. The sensible takeaway is the workflow and its supporting infrastructure, not a claim that 18 lines of intent have abolished software bugs.
What to watch
- Whether Epsilon or AWS publishes test reliability and maintenance results over time.
- How often teams need to correct or review natural-language test steps.
- Whether the approach stays complementary to Playwright as more teams adopt it.
Discussion spark: Should AI-driven tests mainly handle workflows that are awkward to script, or should teams start replacing selector-based tests wherever they can?
Sources and evidence
- Building intent-driven UI testing with Amazon Nova Act (29 October 2026, 18:20 UTC)
- Build a repeatable PHP-to-Java migration engine with Kiro (29 October 2026, 18:20 UTC)
not affiliated with or endorsed by Amazon Web Services (AWS)