Abstract

Browser automation has evolved from simple scripting to sophisticated orchestration frameworks. This paper examines modern approaches to web automation engineering, focusing on deterministic recipe-based execution, vision-augmented fallback systems, and scalable multi-platform deployment.

1. Introduction

The landscape of web automation has shifted dramatically. Modern frameworks like Playwright, Puppeteer, and browser-use provide unprecedented control over browser instances. This work presents a production-grade architecture that combines deterministic recipes with AI-powered autonomous agents.

2. Architecture Overview

A robust automation system requires multiple layers:

• Orchestration Layer: CLI-driven runner that dispatches to appropriate execution engines
• Recipe Engine: JSON-defined step sequences for known platforms (zero LLM cost)
• Agent Layer: Autonomous browser agents using Claude Sonnet for unknown sites
• Vision Fallback: Screenshot-based element detection when DOM selectors fail
• Validation: Post-action screenshot verification with success criteria

3. Recipe-Based Execution

Deterministic recipes encode platform-specific workflows as JSON. This approach eliminates LLM costs for known platforms while maintaining reliability through explicit selector chains.

4. Proxy and Session Management

Residential proxy rotation with sticky sessions ensures consistent IP addresses across multi-step flows. Port-based hashing maps domains to specific proxy endpoints, maintaining session affinity without external state.

5. Verification Framework

Published content must meet strict criteria including HTTP 200 response, no redirects to login pages, proper title tags, no noindex directives, matching H1 headings, and unique domain counting.

6. Conclusion

Browser automation engineering requires balancing determinism with adaptability. Recipe-first approaches minimize cost and maximize reliability, while AI agents handle the long tail of unknown platforms.