Epsilon Scales Intent-Driven UI Testing Across Enterprise Product Teams Using Amazon Nova Act
Global marketing technology leader Epsilon, a subsidiary of Publicis Groupe managing hundreds of millions of consumer profiles and thousands of brand partnerships, has successfully integrated Amazon Nova Act and Amazon Bedrock AgentCore Browser Tool to revolutionize its quality assurance and UI testing infrastructure. Faced with the challenge of keeping pace with rapid application development cycles without imposing heavy maintenance overhead on its internal product teams, Epsilon deployed a multi-tenant, intent-driven testing service as part of its centralized Agentic Platform architecture.
The implementation marks a significant evolution in enterprise software testing methodologies. While traditional selector-based frameworks remain vital for regression suites, Epsilon’s integration of natural-language browser automation bridges the gap between technical execution and business-level intent, enabling product managers and quality assurance leads to author, maintain, and scale complex UI test workflows with unprecedented efficiency.
Background and Enterprise Context
Managing infrastructure for a massive portfolio of digital applications requires balancing rigorous quality control with operational agility. Within Epsilon, the Agentic Platform team—comprising 50 specialized engineers—is tasked with building unified, shared infrastructure to allow diverse product groups to build, deploy, govern, and execute artificial intelligence agents safely. Prior to the adoption of Amazon Nova Act, product teams looking to implement advanced automated testing faced considerable friction, often forced to independently solve complex challenges related to federated identity, pre-flight authorization, Model Context Protocol integration, cross-account provisioning, and underlying browser infrastructure management.
As Epsilon’s digital ecosystem expanded, the limitations of maintaining brittle, selector-dependent testing scripts became increasingly apparent. Minor UI redesigns, DOM structure shifts, or class-name alterations frequently caused test failures, demanding constant code-level maintenance from engineering resources. Recognizing that their existing Playwright-based regression backbone functioned well for core functionality, Epsilon sought an additive solution rather than a replacement. They required a scalable, natural-language testing layer that could operate within a multi-tenant platform framework without requiring individual product teams to configure or manage underlying browser binaries, display servers, or session lifecycles.
The Technological Solution: Architecture and Implementation
To address these enterprise requirements, Epsilon leveraged Amazon Nova Act in combination with the Amazon Bedrock AgentCore Browser Tool, delivering resilient UI testing as a fully managed platform service. The architecture is structured around three primary components: a centralized platform API for workflow creation, a background task execution runner, and the core browser automation engine.
The operational pipeline flows seamlessly from the user interface through an API Gateway equipped with robust authentication and permission checks. User requests are processed by a Django REST API layer that exposes comprehensive endpoints for workflows, executions, automation suites, and system statistics. The business logic layer utilizes a Workflow Manager, Execution Orchestrator, and Suite Coordinator to queue jobs in Amazon Simple Queue Service. These jobs are subsequently processed by a Runner Service operating on Amazon Elastic Container Service, housing the Nova Act engine and a headless Chromium browser instance. All operational results, high-resolution screenshots, trajectories, and comprehensive HTML reports are securely archived in Amazon S3.
This capability is exposed across the organization as a standardized agent designated as the ui-test-agent, equipped with specialized skills. Onboarding new product teams requires minimal administrative effort; the platform team provisions a secure cross-account IAM role, registers the target application URLs, and grants direct access to the testing agent through the pre-existing platform API. Consequently, teams can immediately author workflows and retrieve execution results via familiar interfaces.
Chronology and Implementation Workflow
The deployment of the intent-driven testing framework followed a structured engineering timeline focused on isolation, multi-tenancy, and performance optimization.
In the initial phase, Epsilon’s Agentic Platform team evaluated natural-language authoring tools against strict multi-tenant requirements. Following the selection of Amazon Nova Act and the Amazon Bedrock AgentCore Browser Tool, the engineering group developed a dedicated FastAPI sidecar runner service to isolate execution dependencies and prevent conflicts with existing agent stacks.
A concrete example of this architecture in action is Epsilon’s automated workflow for the platform’s "Create an Agent" feature, which spans 18 distinct interactive steps. The workflow systematically processes authentication, form inputs, conditional UI elements, and validation:
- Inputting the username and proceeding to the next screen.
- Entering credentials and executing verification.
- Navigating to the agent creation interface and selecting specification design cards.
- Dynamically handling onboarding tour pop-ups by selecting skip commands if present.
- Populating designated agent name and description text fields.
- Executing programmatic scroll commands to reveal goal statement parameters.
- Identifying and utilizing dynamic skill search boxes.
- Saving configurations, confirming success prompts, and verifying list integration.
Crucially, steps addressing dynamic interface behaviors—such as bypassing unexpected onboarding tours or scrolling to reveal hidden form fields—demonstrate advanced conditional logic. While traditional automation frameworks demand explicit, complex branching code to handle such UI variations, Amazon Nova Act processes these scenarios naturally through semantic language understanding.
Comparative Efficiency and Supporting Data
Quantitative metrics gathered across Epsilon’s production deployments highlight substantial gains in productivity, resource allocation, and authoring speed when comparing intent-driven testing with traditional methodologies:
| Dimension | Amazon Nova Act | Playwright (Equivalent) |
|---|---|---|
| Lines Authored | 18 natural-language steps | Approximately 80–150 lines of code |
| Selector/DOM Inspection | None required | Required for every interactive step |
| Authoring Time (Per Workflow) | 5 to 10 minutes | 30 to 90 minutes |
| Relative Effort Ratio | Baseline | 4x to 6x greater engineering effort |
Furthermore, Epsilon implemented specific optimization strategies to mitigate operational costs and reduce execution latency. By capturing and reusing recorded trajectories during the draft state, the platform successfully replays validated paths in active production states. This mechanism eliminates redundant large language model processing cycles during routine regression runs, significantly reducing both execution time and cloud compute expenditures.
Industry Implications and Strategic Analysis
The successful deployment of Epsilon’s intent-driven UI testing service signals a broader paradigm shift in enterprise software quality assurance. By abstracting away the mechanical complexities of DOM inspection, element locators, and browser infrastructure management, organizations can democratize quality assurance. Non-technical stakeholders, including product managers, business analysts, and quality assurance specialists, can author and review test suites using plain English, drastically lowering the barrier to entry for comprehensive testing.
Industry analysts note that as software release cycles accelerate to meet modern consumer demands, conventional testing bottlenecks frequently impede deployment frequency. Epsilon’s multi-tenant platform architecture demonstrates how generative AI and managed browser automation can be effectively scaled across large enterprises without compromising security, compliance, or isolation standards. Through the utilization of cross-account IAM role assumptions and managed cloud observability tools like AWS CloudTrail, Epsilon has maintained rigorous auditability while eliminating infrastructure maintenance overhead.
Conclusion and Future Outlook
The integration of Amazon Nova Act and Amazon Bedrock AgentCore Browser Tool has successfully transformed Epsilon’s UI testing operations from a brittle, maintenance-heavy chore into a streamlined, resilient platform capability. By decoupling test definitions from underlying DOM structures, Epsilon has insulated its product teams from the recurring disruptions caused by frequent UI updates.
As enterprises increasingly adopt agentic workflows and decentralized platform models, Epsilon’s implementation serves as a benchmark for scalable, intent-driven automation. The organization plans to continue expanding its platform offerings, applying similar natural-language principles across additional operational domains to further enhance engineering velocity and application reliability across its global network.
Related posts:
