The New Stack Agents – Paul Klein IV, Founder, Browserbase

Browserbase and the Future of AI-Powered Web Automation

Key Concepts:

• AI Agents
• Headless Browsers
• Browser Infrastructure
• Web Automation
• Stagehand (Browserbase's open-source framework)
• Director.ai (Browserbase's application for prototyping AI agents)
• LLMs (Large Language Models)
• VLMs (Vision Language Models)
• Observability
• Deterministic vs. Non-Deterministic Agents

What Browserbase Does

Browserbase helps AI agents interact with the internet by providing browser infrastructure, a tool layer (Stagehand), and an application (Director.ai) for automating web interactions. The goal is to provide a complete toolkit for AI to control websites via a web browser, enabling AI to work on behalf of users.

History of Headless Browsers

Headless browsers originated as a way for developers to test web applications. They simulate a browser environment without a graphical user interface, allowing for automated testing of website functionality. Early examples include PhantomJS, and companies like BrowserStack and Sauce Labs have built businesses around headless browser testing. Frameworks like Selenium, Puppeteer, and Playwright have made it easier to integrate browser control into testing workflows. However, these tools are primarily designed for testing and are brittle by design, meaning they break when website interfaces change.

Browserbase's Approach to Web Automation

Browserbase differentiates itself by focusing on durable automation of work, not just testing. Their framework, Stagehand, is a superset of existing frameworks like Puppeteer and Playwright, but with built-in AI functionality. This allows for more dynamic and resilient automation. Instead of relying on specific code to click a button, Stagehand uses LLMs to identify and interact with elements on a page, even if their appearance or location changes. This enables a single script to control hundreds or thousands of websites.

Example: Instead of specifying "Click the login button, the fifth button on the page, red color," you can tell AI, "Click the login button," and it will figure it out.

Technical Details

• Real Browsers: Browserbase uses real browsers (Chromium) because they are sophisticated networking, rendering, and interaction engines.
• LLMs for Dynamic Functionality: LLMs parse code and determine the appropriate action to take on a page.
• Stagehand: Model-agnostic, open-source framework that integrates with various LLMs.
• Observability: Browserbase provides browser recordings and logging to monitor agent behavior.
• Virtualization: Uses Firecracker virtualization technology.
• Languages: Primarily Go and Node.js.

Use Cases

Browserbase's customers are primarily AI companies and startups building features that act on behalf of users. Examples include:

• Research: Helping people do research on potential sales leads or companies in a specific space (e.g., Clay).
• Workflow Automation: Automating repetitive tasks like file downloads and invoice processing.
• Consumer Applications: Benny, a company that helps people get rebates on food stamps by automating the submission process.
• Procurement: Automating the process of getting quotes and placing orders from multiple vendors.
• Compliance: Automating KYC (Know Your Customer) processes by verifying information on business websites.

Example: A company automating the process of logging into portals, downloading invoices, and uploading them to other portals.

Target User

Browserbase targets developers building AI-powered software and "vibe coders" who use primitives to build personalized software. They aim to empower the next generation of builders by providing the best browser tool for AI applications.

Differentiation and Competition

• Customer Obsession: Browserbase focuses on providing excellent customer support and helping customers succeed.
• Technical Advantages: Observability, model-agnostic framework, and a complete suite of tools.
• AWS Agent Core: While AWS launched a competing product, Agent Core, it is coupled to the Nova Act model, which is not as reliable as other models. Browserbase aims to be cloud-agnostic and provider-agnostic.

The Meeting with AWS

Paul Klein described a meeting with AWS that felt like an attempt to gather information about Browserbase's product. He expressed disappointment with this behavior but emphasized that Browserbase is not worried about the competition.

The Role of LLMs

LLMs enable people to operate with more leverage by automating boring and tedious tasks. They allow users to focus on creative and enjoyable work.

Browsers with Built-In Agents

Browserbase views browsers with built-in agents (e.g., Perplexity's Cosmos) as potential partners. Browserbase can provide the server-side infrastructure for tasks that need to run in the background or when the user is not actively using their device.

The Future of Software

The future of software involves more powerful buttons and hyper-specialized applications. Browserbase aims to be the bridge between AI agents and the legacy internet.

Open Source Strategy

Stagehand is open source to reduce friction and enable more developers to automate the web. Browserbase generates revenue through browser minutes and network gigabytes consumed.

Technical Infrastructure Details

• Chromium: The core browser used by Browserbase.
• Firecracker: Virtualization technology used for security and isolation.
• Go and Node.js: Programming languages used in the infrastructure.

Security Measures

• Observability: Providing customers with full session replays, log lines, and network request information.
• Sandboxing: Isolating browsers and destroying them after use.
• Deterministic Agents: Encouraging customers to break down tasks into atomic steps to improve reliability.

Conclusion

Browserbase is building the infrastructure and tools necessary for AI agents to interact with the web effectively. By focusing on durable automation, providing excellent customer support, and leveraging the power of LLMs, Browserbase aims to be a category-defining company in the AI-powered web automation space.