Skip to main content

Agentic Browser Automation (Bowser Architecture)

Overview​

Agentic browser automation represents a shift in how we approach UI testing and web-based workflows. Rather than relying solely on deterministic scripts, engineers can assemble a reusable stack of skills, subagents, and prompts to offload entire classes of work—such as multi-browser UI validation and complex data extraction—to agents.

The Bowser Architecture is an opinionated, four-layer design for building repeatable and scalable agentic browser automation.

The Four-Layer Architecture​

1. Skill Layer: Foundational Capabilities​

The skill layer defines the low-level capabilities available to an agent. At Solution8, we prioritize technologies that are token-efficient and flexible:

  • Claude + Chrome: Using the --d-chrome flag to inject browser access directly into a Claude instance.
  • Playwright CLI: Preferred over MCP servers for its efficiency and ability to handle parallel headless sessions and persistent login profiles.

2. Agent Layer: Specialized Subagents​

Skills are lifted into specialized agents designed for specific tasks. For example:

  • Browser QA Agent: This agent parses a user story into concrete steps, creates project directories, takes screenshots at each stage, and reports a clear pass/fail status.
  • Playwright Browser Agent: Used for arbitrary, non-interactive automation.
  • Claude Browser Agent: Used for interactive, human-in-the-loop browser control.

3. Orchestration Layer: Team Workflows​

The orchestration layer composes agents and skills into end-to-end workflows using custom commands and reusable prompts.

  • jui review: A command that kicks off multi-browser UI tests by spawning subagents to validate user stories in parallel.
  • Team Orchestration: A primary agent acts as a "manager," spawning subagents as teammates and giving them explicit prompts for specific tasks. Results are then collected, merged, and reported as a unified summary.

4. Reusability Layer: Just Files and Command Runners​

The top layer focuses on making these complex workflows easily accessible and repeatable.

  • just files: We use just as a single entry point to catalog and run preconfigured commands. This exposes workflow permutations and variables for rapid invocation by both engineers and agents.
  • Higher-Order Prompts (hop): A pattern where a prompt accepts another prompt as a parameter, wrapping workflows in a consistent, reusable routine.
  • Automated Workflows: Storing automations (like "Amazon checkout" or "blog aggregation") as files in a dedicated directory allows browser tasks to scale from individual skills into robust, repeatable assets.

Why Agentic UI Testing Scales​

Agentic testing acts like a real user, which offers several advantages over conventional frameworks:

  • Reduced Overhead: Removes much of the configuration and maintenance burden associated with deterministic test suites.
  • Rapid Creation: Enables the quick generation of arbitrary workflows and user stories.
  • Traceability: Agents can produce comprehensive UI summaries with screenshots for every step, ensuring clear audit trails.

Note: An optimal strategy often involves a hybrid of deterministic code tests (for critical logic) and agentic, non-deterministic testing (for broad UI validation).

Risks and Mastery​

While agentic automation provides a decisive advantage, it comes with specific responsibilities:

  • Retaining Control: Engineers must build and understand their own skills and subagents rather than relying solely on external plugins or "borrowed" prompts. This is essential for differentiation and security.
  • Security Risks: Outsourcing agent competence without mastery introduces risks like prompt injection and maintenance debt.
  • Specialization: The ultimate advantage in the age of agents is the ability to combine scale with deep specialization in agent orchestration.