Skip to content
Agents tracked: 492 Releases (7d): 545 Agent status: 1 with issuesSince last limit reset: Claude 9d · Codex 7d For AI agents: Markdown Updated Nov 20, 2026
Case study EpsilonResults as reported by AWS

Epsilon sped up UI test authoring 4-6x with Amazon Nova Act

Epsilon, the Publicis Groupe data and marketing-services company, added plain-English UI tests on top of its Playwright regression suite and offers them to its product teams as a shared platform service. Tests are written as natural-language steps that Amazon Nova Act runs in managed AgentCore browser sessions. Per AWS and Epsilon, a representative 18-step workflow took 5–10 minutes to write instead of 30–90 minutes for the equivalent Playwright script.

The problem

Epsilon's Playwright suite worked well for deterministic regression checks, but scripted tests are tied to selectors and page structure, so cosmetic UI changes can break them even when the feature still works. Several product teams also needed test coverage for their own applications, and setting up browser infrastructure, dependencies and identity integration team by team duplicated effort. The Agentic Platform team (50 engineers) wanted UI testing to be one more managed capability on its agent platform rather than a separate tool every team had to run.

How they did it

  1. Keep your existing scripted suite (Epsilon kept Playwright as its regression backbone) and add intent-based tests on top of it, not instead of it.
  2. Open a managed browser with the AgentCore Browser Tool client, connect Nova Act to it over the Chrome DevTools Protocol WebSocket and authenticate with the normal AWS credential chain instead of separate API keys.
  3. Write each test as a target URL plus a list of plain-English steps, including conditional ones such as "if an onboarding tour box appears, choose Skip Tour".
  4. Run tests from a background worker that takes a screenshot per step, records pass/fail per step and stores results and screenshots in Amazon S3.
  5. Pin Nova Act and the browser client in their own service or container; Epsilon found the Nova Act signing endpoint targets us-east-1 whatever the browser session's region.
  6. Once a workflow passes, save the successful trajectory and replay it for repeat runs, which skips redundant model "thinking" and cuts latency and cost.
  7. For several teams, give each its own scoped cross-account IAM role via AWS STS AssumeRole so every tenant's tests run in its own account.

From the official docs

with NovaAct(cdp_endpoint_url=ws_url, cdp_headers=headers, boto_session=boto_session, starting_page=url) as nova_act: result = nova_act.act(instruction)

Results as reported by AWS

  • Test authoring was 4-6x faster on Epsilon's internal benchmarks, as reported by AWS and Epsilon.
  • A representative 18-step workflow took 5–10 minutes to author with Nova Act against 30–90 minutes for the equivalent Playwright script of 80–150 lines of code, with no selector or DOM inspection, as reported by AWS and Epsilon.
  • Saved trajectories achieved a 100 percent pass rate across repeated runs after a workflow passed the initial explore phase, typically 4–5 prompt iterations, as reported by AWS and Epsilon.
  • Onboarded teams get UI testing through the same platform API as other agents, with no browser binaries, Selenium grid or display servers to manage, as reported by AWS and Epsilon.

These figures are as reported by AWS in AWS Machine Learning Blog customer solution post ↗ (Oct 29, 2026). AWS sells the product used here. AgentGid did not measure or independently verify them.

Takeaway. Natural-language UI tests work best as a layer over a scripted suite, and replaying a saved passing trajectory is what makes them repeatable and cheaper to run.
AgentGid's take

This suits a platform or QA team already on AWS that runs many web apps and wants tests that survive UI redesigns and that product managers can read. It is less of a fit for a small team with one app and a stable Playwright suite, or for anyone outside AWS. Per AgentGid's Amazon Nova Act page (pricing checked Oct 5, 2026), the service costs $4.75 per agent hour, excluding human-in-the-loop wait time, with parallel agents billed separately and availability in US East (N. Virginia); there is no free tier listed, and the Python SDK is open source under Apache-2.0. All figures are self-reported from Epsilon's internal benchmarks in an AWS blog post co-written by AWS and Epsilon staff, not measured independently, and the Playwright time is an estimate for an equivalent script. If you want an open-source option without AWS lock-in, Browser Use and Stagehand in our catalog cover natural-language browser steps on your own infrastructure.

The agent used here

Similar use cases

Guides