Browser Harness — Self-Healing Browser Automation for LLMs

Introduction

Browser Harness is a persistent, self-healing browser runtime built for AI agents. It extends the Browser Use ecosystem with cloud-ready browser instances that LLMs can control to complete complex web tasks, automatically recovering from navigation errors and page changes.

What Browser Harness Does

• Provides persistent browser sessions accessible via Chrome DevTools Protocol
• Implements self-healing navigation that adapts to page layout changes
• Supports multi-tab workflows with isolated session contexts
• Handles authentication persistence across agent sessions
• Offers cloud and VPS deployment modes for remote browser access

Architecture Overview

Browser Harness runs a managed Chromium instance exposed via CDP. A middleware layer intercepts agent commands and applies self-healing logic: if a selector fails, it attempts fuzzy matching, visual element detection, and DOM re-analysis before retrying. Sessions persist across agent reconnections with full cookie and storage state preservation.

Self-Hosting & Configuration

• Deploy via Docker with a single container image
• Configure CDP port and authentication via environment variables
• Set up persistent storage volumes for session state
• Scale horizontally with multiple browser instances behind a load balancer
• Integrate with Telegram or Slack for remote agent monitoring

Key Features

• Self-healing selectors that adapt to DOM changes automatically
• Persistent sessions surviving agent disconnects and reconnects
• Cloud-native deployment with Docker and Kubernetes support
• Built-in screenshot and video recording for debugging
• CDP-compatible interface working with Playwright and Puppeteer

Comparison with Similar Tools

• Browser Use — the Python agent library; Browser Harness is the persistent runtime
• Playwright — browser automation library; Browser Harness adds self-healing and persistence
• Stagehand — AI browser framework; Browser Harness focuses on infrastructure
• Selenium — traditional automation; Browser Harness is designed for LLM agent workflows

FAQ

Q: How does self-healing work? A: When a CSS selector fails, the harness uses fuzzy matching, ARIA attributes, and visual analysis to locate the intended element and retry the action.

Q: Can multiple agents share one browser instance? A: Yes. The harness supports isolated tab contexts so multiple agents can operate concurrently in separate tabs.

Q: What browsers are supported? A: Chromium-based browsers via the Chrome DevTools Protocol.

Q: How do I persist login sessions? A: Session storage and cookies are saved to disk by default and restored on reconnect.

Sources

• https://github.com/browser-use/browser-harness