Architecture

Building Beryl: Why We Built a Crawler that Truly Understands Auth

November 10, 2026

The Great Wall of Authentication

When we first sat down to build Beryl, we were obsessed with the idea of autonomous AI testing. We wanted an agent that could navigate an application, find bugs, and report back, all without human intervention. We had the computer vision models. We had the reasoning engines. We felt invincible.

And then we hit our first real-world customer, and our invincible AI agent slammed face-first into a brick wall: the login screen.

We realized very quickly that 90% of the modern web—and 100% of the valuable parts—is hidden behind authentication.

The Duct-Tape Era of Testing

Historically, the software industry has treated authentication in testing as a massive inconvenience to be bypassed rather than a core feature to be tested. Engineers would use 'duct-tape' solutions. They would write brittle scripts to inject session cookies directly into the browser. They would turn off Multi-Factor Authentication in staging. They would hardcode fake admin tokens into the backend just so the automated tests could get past the login screen.

This approach creates a massive blind spot. If your testing suite actively avoids logging in the way a real user logs in, you are completely blind to authentication failures. When a third-party SSO provider updates their flow, or when a session token expiry bug hits production, your tests will stay green while your customers are completely locked out of the product.

The AI Reality Check

As we built Beryl, we knew we couldn't rely on these hacky bypasses. If an AI agent cannot independently log into an application, it cannot test it. If it gets confused by a CAPTCHA, or stuck in a Google OAuth redirect loop, the entire autonomous QA pipeline grinds to a halt.

We had to stop treating authentication as an obstacle and start treating it as a first-class citizen of the crawler's architecture.

Building the Master Key

We spent months building an entirely new kind of crawler. We didn't build a web scraper; we built an autonomous entity that natively understands authentication state.

When Beryl's crawler encounters a login gate today, it doesn't crash or ask for a magic token. It visually identifies the username and password fields. It securely pulls the correct credentials from an encrypted vault. It types them in exactly like a human would. If it encounters a multi-factor authentication prompt or a CAPTCHA, it solves it autonomously.

By building a crawler that truly understands auth, we gave our AI the master key to the application. It doesn't bypass reality; it navigates it. And because it tests the real login flow every single time, our customers sleep soundly knowing that the front door to their business is always open.

Author

Tanay Agarwal

Tanay Agarwal

Software Engineer